NOUN Accept All Cookie Settings https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 1/21

About OED How to use the OED https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 20/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary Historical Thesaurus Purchasing Editorial policy Help with access Updates World Englishes Institutional account management Contribute Accessibility Oxford University Press Contact us Oxford Languages Upcoming events Oxford Academic Case studies Oxford Dictionary of National Biography Media enquiries Oxford University Press is a department of the University of Oxford. It furthers the University's objective of excellence in research, scholarship, and education by publishing worldwide Cookie policy Privacy policy Legal notice Copyright © 2024 Oxford University Press

https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 21/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary sports agent, n. 1943– A person who represents a professional athlete in sleeper agent, n. 1945– = sleeper, n. I.2d. bioagent, n. 1950– A harmful or disease-producing microorganism, biopesticide, biotoxin, etc., esp. one used in warfare or for the purposes of terrorism. G-agent, n. 1953– Any of a group of four organophosphorus nerve agents originally developed by German scientists during the Second World War, characterized by being… uncoupling agent, n. 1956– = uncoupler, n. stripping agent, n. 1958– nerve agent, n. 1960– A substance that alters the functioning of the nervous system, typically inhibiting neurotransmission; esp. one used as a weapon, a nerve gas. Agent Orange, n. 1966– A defoliant and herbicide used by the United States during the Vietnam War to remove forest cover and destroy crops. Cf. agent, n.¹ & adj.compounds… penetration agent, n. 1966– A spy sent to penetrate an enemy organization. treble agent, n. 1967– A spy who works for three countries, his or her superiors in each being informed of his or her service to the other, but usually with actual… triple agent, n. 1968– = treble agent, n. managing agent, n. 1969– A person responsible for administering or managing an activity (esp. a sale) on behalf of another; (Insurance) a manager of an underwriting syndicate… masking agent, n. 1977– A chemical compound which conceals the presence of a substance within the body; (Sport) a

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary A substance, such as yeast or baking powder, which is used in dough or batter to make it rise during (and sometimes before) baking. travel agent, n. 1885– A person who owns or works for a travel agency; (also) a travel agency. polling agent, n. 1887– An ofstation on the day of an election. special agent, n. 1893– A person who conducts investigations on behalf of the government; (now) spec. (U.S.) a person who conducts criminal investigations and has arrest… transport-agent, n. 1897– alkylating agent, n. 1900– A substance that brings about alkylation; (Pharmacology) any of a class of cytotoxic immunosuppressant drugs which alkylate DNA and are used in… addition agent, n. 1909– (In electrodeposition) a substance which is added to an electrolyte, typically in small quantities, in order to modify the quality of the deposit… site agent, n. 1910– a. An agent authorized to inspect, survey, and purchase land for development (rare); b. (in the construction industry) a person responsible for… marketing agent, n. 1915– harassing agent, n. 1919– A non-lethal chemical which is deployed in the form of a gas or aerosol and used to incapacitate an enemy or disperse a crowd; = harassing gas, n. contrast agent, n. 1924– A substance introduced into a part of the body to enhance the quality of a radiographic image by increasing the contrast of internal structures with… binding agent, n. 1933– A substance that assists cohesion (cf. bind, v. III.10). stock-agent, n. 1933– Oxfodrodu Ubnleiv aergseitnyt P, rne.ss u1s9e3s5 c–ookies to enhance your experience on our website. By selecting ‘accept all’ you are Aa gsrpeye iwngh oto w oourrk uss oen o fb ceohoaklfie osf. Ymouut ucaanll yc hhaonsgteil ey ocuoru cnotorikeise, suestutianlglys awti athn ya ctitmuea.l Malolergei ainnfcoerm oantliyo nto can be founodn ein. our Cookie Policy. release agent, n. 1938– A substance which is applied to a surface in order to prevent adhesion to it. https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 19/21

tourist agent, n. 1884– raising agent, n. 1885– https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 18/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary booking agent, n. 1849– a. A person who or a business which arranges transport or travel for goods or passengers, or sells tickets in advance for concerts, plays, or other… Frequently derogatory in early use, denoting agents for railway or shipping companies who issued tickets or passes which were greatly overpriced or invalid; cf. booker, n. 3a. passenger agent, n. 1852– baggage-agent, n. 1858– employment agent, n. 1859– An individual acting as a professional intermediary between applicants for work and employers. claim-agent, n. 1860– matrimonial agent, n. 1860– personation agent, n. 1864– An ofadvance agent, n. 1865– An agent who is sent on ahead of a main party (cf. advance man, n.); also transfer agent, n. 1869– information agent, n. 1871– mission-agent, n. 1871– lecture agent, n. 1873– publicity agent, n. 1877– agent word, n. 1879– A word that indicates agency or active force; esp. a word that denotes the doer of an action; = agent noun, n. personating agent, n. 1879– = personation agent, n. rental agent, n. 1880– bittering agent, n. 1883–

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary business agent, n. 1831– customs agent, n. 1838– = customs ofmine agent, n. 1839– agentive, adj. & n. 1840– Of or relating to an agent or agency (see agent, n.¹ A.1c); indicating or having the semantic role of an agent. railroad agent, n. 1840– road agent, n. 1840– †a. An agent or driver for a stagecoach company (obsolete); b. a robber who steals from travellers or holds up vehicles on the road (now historical). rogue agent, n. 1840– station agent, n. 1840– a. Chieworks for (a particular branch of) an intelligence… A substance used to clarify a liquid; spec. (a) a substance used to remove organic compounds from a liquid, esp. beer or wine, to improve the clarity… freight agent, n. 1843– shipping-agent, n. 1843– A licensed agent who transacts a ship's business for the owner. goods agent, n. 1844– intelligence agent, n. 1844– patent agent, n. 1845– land-agent, n. 1846– A steward or manager of landed property; also, an agent for the sale of land, an estate agent. change agent, n. 1847– A person who initiates social or political change within a group or institution. Oxfo p r a d y U a n g i e ve n r t s , i t n y . Pre1s8s4 u7s–es cookies to enhance your experience on our website. By selecting ‘accept all’ you are Aangr oeeffounreds ipno onusri b Cloeo fkoire a Pdovliicsyin. g the U.S. president on rates of… bureau agent, n. 1848– An agent or ofhttps://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 17/21

agentless, adj. 1831– Lacking an agent (in various senses); without an agent. https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 16/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary An agent (now typically a professional one) who acts on behalf of an author in dealing with publishers and others involved in promoting his or her… theatrical agent, n. 1797– An agent whose business is to act as an intermediary between actors looking for work and those seeking to employ them. commission agent, n. 1798– †a. = commission broker, n. (a) (obsolete); b. an agent who conducts business or trade for another party on the principle of commission (commission… book agent, n. 1810– A person who promotes the sale of books; (now) spec. a literary agent (cf. agent, n.¹ A.2e). forwarding agent, n. 1810– A person or business that organizes the shipment or transportation of goods. newsagent, n. 1811– A dealer in newspapers and periodicals, esp. the owner of a shop where these are sold; (now also) the shop itself, usually also selling tobacco… police agent, n. 1813– ship-agent, n. 1813 A shipping agent. oxidizing agent, n. 1814– A substance that brings about oxidation and in the process is itself reduced. press agent, n. 1814– A person employed to organize advertising and publicity in the press on behalf of an organization or person. reducing agent, n. 1816– A substance that brings about chemical reduction and in the process is itself oxidized; cf. oxidizing agent, n. parliamentary agent, n. 1819– A person professionally employed to take charge of the interests of a party concerned in or affected by any private legislation. counter-agent, n. 1821– A counteracting agent or force; a counteractant.

s – cookies to enhance your experience on our website. By selecting ‘accept all’ you are agreeing to our use of cookies. You can change your cookie settings at any time. More information can be An agent employed (by the landlord or owner) in letting or selling a house, collecting rents, etc.; found in our Cookie Policy. (now esp.) an estate agent. literary agent, n. 1794– https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 15/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary crown agent, n. 1753– An agent for the Crown; spec. (usually with capital initials) (a) in Scotland, a law ofcharge of criminal proceedings, acting under… agentess, n. 1757– A female agent. navy agent, n. 1765– A person or paymaster or purser in the U.S. navy (obsolete). Indian agent, n. 1766– An ofpeople; (in Canada) the chief government… prize agent, n. 1766– An agent appointed for the sale of prizes taken in maritime war. An agent (now esp. an intelligence agent) who works away from a central ofadvertising agent, n. 1775– purchasing agent, n. 1777– coal agent, n. 1778– federal agent, n. 1781– A representative of the U.S. federal government, (now) esp. a federal law-enforcement ofnewspaper agent, n. 1781– agent noun, n. 1782– A noun (in English typically one ending in -er or -or) denoting someone or something that performs the action of a verb, as worker, accelerator, etc. estate agent, n. 1787– A person or company involved in the business or profession of arranging the sale, purchase, or rental of buildings and land for clients. Also (also… revenue agent, n. 1787– recruiting agent, n. 1792– Oxfohrodu Useni vaegresnityt ,P nre.ss u

agentship, n. 1608– The position, role, or function of an agent (in various senses); agency. Also: an instance of this. https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 13/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary Frequency of agent, n.¹ & adj., 2017–2023 * Occurrences per million words in written English Modern frequency series are derived from a corpus of 20 billion words, covering the period from 2017 to the present. The corpus is mainly compiled from online news sources, and covers all major varieties of World English. Compounds & derived words Sort by Date (oldest nihilagent, n. 1579–80 A person who does nothing. agentry, n. 1590– The ofagency; the process or fact of being an agent… vice-agent, n. 1597–

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary Frequency agent is one of the 1,000 most common words in modern written English. It is similar in frequency to words like agree, distribution, kill, military, and sell. It typically occurs about 100 times per million words in modern written English. agent is in frequency band 7, which contains words occurring between 100 and 1,000 times per million words in modern written English. More about OED's frequency bands Frequency data is computed programmatically, and should be regarded as an estimate. Frequency of agent, n.¹ & adj., 1750–2010 * Occurrences per million words in written English Historical frequency series are derived from Google Books Ngrams (version 2), a data set based on the Google Books corpus of several million books printed in English between 1500 and 2010. The overall frequency for a given word is calculated by summing frequencies for the main form of the word, any plural or inFor sets of homographs (distinct entries that share the same word-form, e.g. mole, n.¹, mole, n.², mole, n.³, etc.), we have estimated the frequency of each homograph entry as a fraction of the total Ngrams frequency Oxfofordr t Uhen iwveorrsdi-tfyo Prmre. s Tsh uiss mesa yc oreoskuielts i nto i neanchcuarnacceie yso.ur experience on our website. By selecting ‘accept all’ you are agreeing to our use of cookies. You can change your cookie settings at any time. More information can be found in our Cookie Policy. https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 12/21

1600s agentt https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 11/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary 1535 The fynall necessytie also, and the cause agent [Latin causam agentem] or eW. Marshall, translation of Marsilius of Padua, Defence of Peacei. viii. f. 67v 1575 The ayre being more thin and liquide then the water, and more vnable to resist, is sooner and more easily atranslation of L. Daneau, Dialogue Witches iii. sig. E.vii 1615 Hughe Mill and Elinor his wife the parties agentes in this cause and William delve defendent. in B. Cusack, Everyday English 1500–1700 (1998) 24 1620 What a hot fellow Sol (whom all Agent Causes follow). J. Melton, Astrologaster 13 1704 The proper oJ. Norris, Essay Ideal Worldvol. II. vii. 350 1856 Agent or patient, singly or one of a crowd. T. De Quincey, Confessions Eng. Opium-eater (revised edition) in Selections Grave & Gayvol. V. 83 1949 The Philosopher is speaking in that passage not of the agent cause but of the formal cause. M. C. Fitzpatrick, translation of St. Thomas Aquinas, On Spiritual Creatures i. 25 2009 The [Philippine] people have transmogriagent force of revolution. N. X. M. Tadiar, Things fall Away vii. 290 Pronunciation BRITISH ENGLISH U.S. ENGLISH /ˈeɪdʒ(ə)nt/ /ˈeɪdʒ(ə)nt/ AY-juhnt AY-juhnt Pronunciation keys Forms Variant forms late Middle English– agent

Acting, exerting power (sometimes contrasted with patientadj. A.2a). 1535– † party agentnoun Obsolete Law the person or party bringing a suit. https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 10/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary Sometimes overlapping with sense A.1b. 1579 The gallowes is no agent or doer in those good thinges. W. Fulke, Heskins Parleament Repealed in D. Heskins Ouerthrowne 621 1593 Not a nayle in it [sc. the Crosse] but is a necessary Agent in the Worlds redemption. T. Nashe, Christs Teares 21/1 a1616 Here is her hand, the agent of her heart. W. Shakespeare, Two Gentlemen of Verona (1623) i. iii. 46 1654 God doth often good works by ill agents. J. Bramhall, Just Vindication of Church of England iii. 43 1793 War, which is the agent which must in general be employed upon these occasions, presents..an uncertain court of judicature. B. Vaughan, Letters Concert of Princes p. iii 1842 Nature..Thro' many agents making strong, Matures the individual form. Lord Tennyson, Love thou thy Land in Poems (new edition) vol. I. 225 1878 Whatever thus furnishes us with the that is, something which acts for us and assists us. W. S. Jevons, Political Economy 26 1920 Money is the agent through which good purposes are made eIntellectvol. 12 233/2 2002 [In Marlowe's physiology] the arteries..carry the vital spirit..which is the agent by which the soul eY. Takahashi in S. W. Wells, Shakespeare Surv. 181 4. Chemistry. A substance that brings about a chemical or physical effect or causes a chemical 1624– reaction. In later use chieeffect or reaction. Cf. reagentn. 2. alkylating, oxidizing, reducing, wetting agent, etc.: see the 1624 The vinegre..is the onely Agent [French l'vnique agent; Latin solum medium aptum] in the whole World for this Art, that can resolue and reincrudate, or make raw againe the Mettallicke Bodies. ‘E. Orandus’, translation of N. Flamel, Expos. Hieroglyphicall Figures St. Innocent's Church-yard 159 1671 The agent in the change wrought by Petrisaxeous odour, or invisible ferment.

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary mis-agent, n. 1625 non-agent, n. 1632– smock-agent, n. 1632– agent, v. 1637– transitive. To act as agent in (some business or process); to conduct or carry out as agent. Also: to act as an agent for (a person or project). agenting, n. 1646– The business or process of acting as an agent (in various senses); the profession of an agent. foreign agent, n. 1646– A person who represents or acts on behalf of one country while located in another; (in later use spec.) a person who works secretly to obtain… free agent, n. 1649– a. A person able to act freely, as by the exercise of free will, or because of the absence of restriction, constraint, or responsibilities; b. Sport… reagent, n. 1656– Chemistry. A substance used in testing for other substances, or for reacting with them in a particular way; (more widely) any substance used in… agent general, n. 1659– spec. (sometimes with capital initials). Formerly: the representative of a British colony in London (now historical). Later: the representative of an… under-agent, n. 1677– A sub-agent. subagent, n. 1683– A subordinate agent; (U.S. Law) an agent authorized to transact business or otherwise act on behalf of another. chemical agent, n. 1728– A chemical substance producing a speci(now often) spec. a substance used to incapacitate… inter-agent, n. 1728– An intermediate agent; a go-between, intermediary. Oxfocordm Umnievrecrsiaitly a Pgreesnst ,u nse.s co1o73k7ie–s to enhance your experience on our website. By selecting ‘accept all’ you are Aa gpreeresinogn toor o ourrg uasneiz oaft cioono kaiuesth. Yooriuz ecadn t och aacntg oen y oaunro tchoeork'ise bseehttainlfg isn a mt aanttye trims ree. l Matoinreg itnofo crommamtioenr ccean be found in our Cookie Policy. or trade; spec. (U.S.) an oftravelling agent, n. 1737– A travelling salesperson; a representative who travels on behalf of a company. https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 14/21

3. The means by which something is done; the material cause or instrument through which an 1579– effect is produced (often implying a rational employer or contriver). https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 8/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary 1825 Mr. Schemer, the agent, had no situation for our hero upon his books, but Proteus heard..that Mr. Make-a-bill..was in great want of a person at his theatre. P. Egan, Life of Actor vi. 220 1892 By an early hour of the numbered evening I might have been observed..dining with my agent. R. L. Stevenson & L. Osbourne, Wrecker vi. 95 1917 The name on the door was Abe Riesbitter, Vaudeville Agent, and from the other side of the door came the sound of many voices. P. G. Wodehouse, Man with Two Left Feet 34 1946 Mr Watt, my agent, and Mr Faber, my publisher, have Daimlers and country cottages. P. Larkin, Letter 28 July in Selected Letters (1992) 120 1970 Her agent..was nonplussed. ‘Look, baby,’ he gently chided, ‘we're walking away with one million..dollars a picture.’ T. Southern, Blue Moviei. viii. 64 1983 Agents, often to justify their percentage when all they really do for a big star is make a phone call, are geniuses when it comes to new things to ask for. W. Goldman, Adventures in Screen Trade 18 2003 [She] was no longer the timid, inexperienced ingénue..protected by her agent. C. Fitheatre 2.f. U.S. A stagecoach robber; = road agentn. Now historical. 1876– 1876 The driver San Andreas. Weekly Calaveras Chronicle (Mokelumne Hill, California) 29 July 3/1 1880 We reached it before long, and concluded that the ‘agents’, or robbers, had an excellent eye for position. A. A. Hayes, New Colorado (1881) xi. 154 1904 Nex' time I drives stage some of these yere agents massacrees me from behind a bush. S. E. White, Blazed Trail Storiesii. iii. 155 1970 The agents developed a system of marking departing stagecoaches that were carrying treasure so that confederates would know which ones to stop. H. S. Drago, Great Range Wars xviii. 207

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary 1571 Faieth is produced and brought foorth by the grace of God, as chiefe agent and worker thereof. W. Fulke, Confut. Popishe Libelle (new edition) f. 108v 1592 I stepped back againe into the garden,..leauing them still agents of these vnkind villanies. R. Greene, Philomela sig. F4v 1645 The distans. A. Ross, Philosophicall Touch-stone 35 1666 Whether or no the Shape can by Physical Agents be alter'd.., yet mentally both..can be done. R. Boyle, Origine of Formes & Qualities 9 1699 When the Samians invaded Zancle, a..great Agent in that aR. Bentley, Dissertation upon Epistles of Phalaris (new edition) 155 1719 I was still to be the wilful Agent of all my own Miseries. D. Defoe, Life Robinson Crusoe 43 1722 Nor can I think, that any body has such an idea of chance, as to make it an agent or really existing and acting cause of any thing. W. Wollaston, Religion of Nature v. 60 1848 Successful production..depends more on the qualities of the human agents, than on the circumstances in which they work. J. S. Mill, Principles of Political Economyvol. I.i. vii. 123 1875 The Rhizopods were important agents in the accumulation of beds of limestone. J. W. Dawson, Life's Dawn on Earth vi. 134 1904 The glacier will be eJournal of Geology (Chicago) vol. 12 574 1963 The key idea of man as the agent for the whole future of evolution. J. S. Huxley, Human Crisis 19 2010 At Cambridge..I had no theories about theatre as an agent of social or political change. S. Fry, Fry Chronicles 94 1.c. Grammar. The doer of an action, typically expressed as the subject of an active verb or in c1620– a by-phrase with a passive verb. Cf. agent nounn. c1620 The active verb adheres to the person of the agent; As, Christ hath conquered hel and

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary agent NOUN1 & ADJECTIVE Etymology Summary Of multiple origins. Partly a borrowing from French. Partly a borrowing from Latin. Etymons:Frenchagent; Latinagent-, agē ns, agere. < (i) Middle Frenchagent (Frenchagent) (noun) person acting on behalf of another, representative, emissary (1332 in an isolated attestation, subsequently (apparently after Italian) from 1578), person who or thing which acts upon someone or something (c1370, originally and frequently in philosophical contexts), substance that brings about a chemical effect or causes a chemical reaction (1612 (in the passage translated in quot. 1624 at sense A.4) or earlier; rare before early 19th cent.), person who intrigues (1640), (adjective) that acts, that exerts power (1337; c1450 in grammar; second half of the 15th cent. in cause agent (compare quot. 1535 at sense B)), and its etymon (ii) classical Latinagent-, agē ns acting, active, (masculine noun) pleader, advocate, in post-classical Latin also representative, ofchurch (6th cent.), (neuter noun) (in philosophy) instrumentality, cause (from 8th cent. in British sources; also in continental sources), uses as adjective and noun of present participle of agere to act, do (see actv.). With sense A.1a and corresponding adjectival use compare earlier patientn. and patientadj. Notes Parallels in other European languages. Compare Catalanagent, adjective and noun (14th cent.), Spanishagente (late 14th cent. as noun, early 15th cent. as adjective), Portugueseagente, adjective and noun (15th cent.), Italianagente (a1294 as adjective, a1328 as noun). Compare also Dutchagent (noun) of(masculine noun) representative, emissary (1546), spy (18th cent., now the usual sense), Agens (neuter noun) person who or thing which acts upon someone or something (1598).

matters for an actor, performer, writer, etc. In earliest use: a theatrical agent. literary, press, publicity, sports agent, etc.: see the https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 7/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary 1764 ReC. Wiseman, Compl. English Grammar 155 1771 An active verb..necessarily supposes an agent, and an object acted upon; as..I praise John. D. Fenning, New Gram. English Tongue 32 1845 It often becomes necessary to state the object of a verb active, or the agent of a verb passive. Hence arises the necessity for..the accusative and the ablative. Encyclopædia Metropolitana (1847) vol. I. 33/1 1953 With an intransitive verb the subject is as much a patient as an agent. I walk is as much ‘I cause my walking’ as ‘I experience my walking’. W. J. Entwhistle, Aspects of Language vi. 179 2007 Truck driver is an acceptable (and existing) compound..but child-driver is not acceptable..since child is the agent of the verb. N. Tsujimura, Introd. Japanese Linguisticsiv. vii. 166 grammar 1.d. Parapsychology. In telepathy: the person who originates an impression (opposed to the 1883– percipient who receives it). 1883 In Thought-transference..both parties (whom, for convenience' sake, we will call the Agent and the Percipient) are supposed to be in a normal state. Proceedings of Society for Psychical Research 1882–3vol. 1 119 1886 We call the owner of the impressing mind the agent, and the owner of the impressed mind the percipient. E. Gurney et al., Phantasms of Livingvol. I. 6 1961 Spontaneous cases [of telepathy] do occasionally occur in which no such connection between apparent agent and apparent percipient can be traced. W. H. Salter, Zoar xi. 149 1990 Analytical attention..has shifted down the years from agent (sender) to percipient (receiver). L. Picknett, Encycl. Paranormal 218/1 parapsychology 2. A person acting on behalf of another. 2.a. A person who acts as a substitute for another; one who undertakes negotiations or 1523– transactions on behalf of a superior, employer, or principal; a deputy, steward,

1652 John and Peter (1 The Agent.) travelled together to (2 The Verb.) Rome. F. Lodowyck, Ground-work New Perfect Language 15 https://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 3/21

cause of some process or change. Frequently with for, in, of. Sometimes dihttps://www.oed.com/dictionary/agent_n1?tab=meaning_and_use#8694696 2/21

10/9/24, 7:26 PM agent, n.¹ & adj. meanings, etymology and more | Oxford English Dictionary 1.a. A person who or thing which acts upon someone or something; one who or that which a1500– exerts power; the doer of an action. Sometimes contrasted with the patient (instrument, etc.) undergoing the action. Cf. actorn. 3a. Earliest in Alchemy: a force capable of acting upon matter, an active principle. Now chiesociological contexts. a1500 The fyrst [kind of combining] is callyd by phylosophers dyptatyve be-twyxte ye agent & ye (1471) pacyent. G. Ripley, Compend of Alchemy (Ashmole MS.) l. 718 a1555 The forgeuenes of oure sinnes..is onely gods worke & we nothing els but patientes & not agentes. J. Bradford, Godlie Medit. Lordes Prayer (1562) sig. Q.ii 1614 For he maketh foure originals, whereof three are agents, and the last passiue and materiall. W. Raleigh, History of Worldi.i. i. §6. 6 1646 Nor are we to be meer instruments moved by the will of those in authority..but are morall Agents. S. Bolton, Arraignment of Errour 295 1788 He that is not free is not an agent, but a patient. J. Wesley, Serm. Several Occasionsvol. V. 177 1809 Agent and Patient, when the same person is the doer of a thing, and the party to whom done: as where a woman endows herself of the best part of her husband's possessions. T. E. Tomlins, Jacob's Law-dictionary 1870 In conformity with this view, the distinction between agent and patient, between something which acts and some other thing which is acted upon, is formally abolished. F. C. Bowen, Logic xii. 401 1909 We are..conversant with the fact in human ais an intelligent agent. Popular Science Monthly April 379 1989 It is silly to berate the hurricane for irresponsibility... It..cannot be a true agent; it cannot author or own an action. C. T. Sistare, Responsibility & Criminal Liability ii. iv. 15 2010 It is only an exercise of power if the agent gets the subject to do something whether or not the subject wants to do it. J. R. Searle, Making Social World vii. 152

AGENT, Black's Law Dictionary (12th ed. 2024) - vice-commercial agent (1800) Hist. In the consular service of the United States, a consular officer who was substituted temporarily to fill the place of a commercial agent who was absent or had been relieved from duty. Westlaw. © 2024 Thomson Reuters. No Claim to Orig. U.S. Govt. Works. End of Document © 2024 Thomson Reuters. No claim to original U.S. Government Works. © 2024 Thomson Reuters. No claim to original U.S. Government Works. 5

AGENT, Black's Law Dictionary (12th ed. 2024) - showing agent (1901) A real-estate broker's representative who markets property to a prospective purchaser. • A showing agent may be characterized as a subagent of the listing broker, as an agent who represents the purchaser, or as an intermediary who owes an agent's duties to neither seller nor buyer. — Also termed selling agent. Cf. listing agent. - soliciting agent (1855) 1. Insurance. An agent with authority relating to the solicitation or submission of applications to an insurance company but usu. without authority to bind the insurer, as by accepting the applications on behalf of the company. 2. An agent who solicits orders for goods or services for a principal. 3. A managing agent of a corporation for purposes of service of process. - special agent (17c) 1. An agent employed to conduct a particular transaction or to perform a specified act. Cf. general agent (1). 2. Insurance. An agent whose powers are usu. confined to soliciting applications for insurance, taking initial premiums, and delivering policies when issued. — Also termed (in sense 2) local agent; solicitor. - specially accredited agent (1888) An agent that the principal has specially invited a third party to deal with, in an implication that the third party will be notified if the agent's authority is altered or revoked. - statutory agent (1844) An agent designated by law to receive litigation documents and other legal notices for a nonresident corporation. • In most states, the secretary of state is the statutory agent for such corporations. Cf. agency by operation of law (1) under agency (1). - stock-transfer agent (1873) See transfer agent. - subagent (18c) 1. A person to whom an agent has delegated the performance of an act for the principal; a person designated by an agent to perform some duty relating to the agency. • If the principal consents to a primary agent's employment of a subagent, the subagent owes fiduciary duties to the principal, and the principal is liable for the subagent's acts. — Also termed subservant. Cf. primary agent; subordinate agent. “By delegation … the agent is permitted to use agents of his own in performing the function he is employed to perform for his principal, delegating to them the discretion which normally he would be expected to exercise personally. These agents are known as subagents to indicate that they are the agent's agents and not the agents of the principal. Normally (though of course not necessarily) they are paid by the agent. The agent is liable to the principal for any injury done him by the misbehavior of the agent's subagents.” Floyd R. Mechem, Outlines of the Law of Agency § 79, at 51 (Philip Mechem ed., 4th ed. 1952). 2. See buyer's broker under broker. - subordinate agent (17c) An agent who acts subject to the direction of a superior agent. • Subordinate and superior agents are coagents of a common principal. See superior agent. Cf. subagent (1). - successor agent (1934) An agent who is appointed by a principal to act in a primary agent's stead if the primary agent is unable or unwilling to perform. - superior agent (17c) 1. An agent on whom a principal confers the right to direct a subordinate agent. See subordinate agent. 2. See high-managerial agent (1). - transfer agent (1850) An organization (such as a bank or trust company) that handles transfers of shares for a publicly held corporation by issuing new certificates and overseeing the cancellation of old ones and that usu. also maintains the record of shareholders for the corporation and mails dividend checks. • Generally, a transfer agent ensures that certificates submitted for transfer are properly indorsed and that the transfer right is appropriately documented. — Also termed stock-transfer agent. - trustee-agent A trustee who is subject to the control of the settlor or one or more beneficiaries of a trust. See trustee (1). - undercover agent (1930) 1. An agent who does not disclose his or her role as an agent. 2. A police officer who gathers evidence of criminal activity without disclosing his or her identity to the suspect. - underwriting agent (1905) Insurance. 1. An agent who acts on behalf of an insurance company to provide insurance to a customer. — Also termed policywriting agent. 2. An agent who acts for an individual Lloyd's underwriter and manages the underwriting syndicate of which the underwriter is a member. — Also termed managing agent. See lloyd's underwriters. 3. An agent who acts for an individual Lloyd's underwriter in all respects except for managing the underwriting syndicate. — Also termed (in sense 3) member's agent. See lloyd's underwriters. - undisclosed agent (1863) An agent who deals with a third party who has no knowledge that the agent is acting on a principal's behalf. Cf. undisclosed principal under principal (1). - universal agent (18c) An agent authorized to perform all acts that the principal could personally perform. © 2024 Thomson Reuters. No claim to original U.S. Government Works. 4

AGENT, Black's Law Dictionary (12th ed. 2024) - land agent See land agent. - listing agent (1927) The real-estate broker's representative who obtains a listing agreement with the owner. Cf. selling agent; showing agent. - local agent (1804) 1. An agent appointed to act as another's (esp. a company's) representative and to transact business within a specified district. 2. See special agent (2). - managing agent (1812) 1. A person with general power involving the exercise of judgment and discretion, as opposed to an ordinary agent who acts under the direction and control of the principal. — Also termed business agent. 2. See underwriting agent (2). - managing general agent (1954) Insurance. A wholesale insurance intermediary who is vested with underwriting authority from an insurer. • Managing general agents allow small insurers to purchase underwriting expertise. They typically become involved in policies that require specialized expertise, as with those for professional liability. — Abbr. MGA. - member's agent See underwriting agent (3). - mercantile agent (18c) An agent employed to sell goods or merchandise on behalf of the principal. — Also termed commercial agent. - nonservant agent (1920) An agent who agrees to act on the principal's behalf but is not subject to the principal's control over how the task is performed. • A principal is not liable for the physical torts of a nonservant agent. See independent contractor. Cf. independent agent; servant. - ostensible agent See apparent agent. - patent agent (1859) A specialized legal professional — not necessarily a lawyer — who has fulfilled the U.S. Patent and Trademark Office requirements as a representative and is registered to prepare and prosecute patent applications before the PTO. • To be registered to practice before the PTO, a candidate must establish mastery of the relevant technology (by holding a specified technical degree or equivalent training) in order to advise and assist patent applicants. The candidate must also pass a written examination (the “Patent Bar”) that tests knowledge of patent law and PTO procedure. — Often shortened to agent. — Also termed registered patent agent; patent solicitor. Cf. patent attorney. - policywriting agent See underwriting agent (1). - primary agent (18c) An agent who is directly authorized by a principal. • A primary agent generally may hire a subagent to perform all or part of the agency. Cf. subagent (1). - private agent (17c) An agent acting for an individual in that person's private affairs. - process agent (1886) A person authorized to accept service of process on behalf of another. See registered agent. - procuring agent (1954) Someone who obtains drugs on behalf of another person and delivers the drugs to that person. • In criminal-defense theory, the procuring agent does not sell, barter, exchange, or make a gift of the drugs to the other person because the drugs already belong to that person, who merely employs the agent to pick up and deliver them. - public agent (17c) A person appointed to act for the public in matters relating to governmental administration or public business. - real-estate agent (1844) An agent who represents a buyer or seller (or both, with proper disclosures) in the sale or lease of real property. • A real-estate agent can be either a broker (whose principal is a buyer or seller) or a salesperson (whose principal is a broker). — Also termed estate agent. Cf. realtor; real-estate broker under broker. - record agent See insurance agent. - registered agent (1809) A person authorized to accept service of process for another person, esp. a foreign corporation, in a particular jurisdiction. — Also termed resident agent. See process agent. - registered patent agent See patent agent. - resident agent See registered agent. - secret agent See secret agent. - self-appointed agent (18c) 1. Someone who is not authorized to act on behalf of another person or entity but who behaves as if such authority has been granted. 2. An agent appointed directly by a principal who also has a statutory agent. 3. A plaintiff in a class-action lawsuit who purports to represent the entire class. - selling agent (1839) 1. The real-estate broker's representative who sells the property, as opposed to the agent who lists the property for sale. 2. See showing agent. Cf. listing agent. - settlement agent (1952) See closing agent. © 2024 Thomson Reuters. No claim to original U.S. Government Works. 3

AGENT, Black's Law Dictionary (12th ed. 2024) - commission agent (1812) An agent whose remuneration is based at least in part on commissions, or percentages of actual sales. • Commission agents typically work as middlemen between sellers and buyers. — Also termed commercial agent. - common agent (17c) An agent who acts on behalf of more than one principal in a transaction. Cf. coagent. - corporate agent (1819) An agent authorized to act on behalf of a corporation; broadly, all employees and officers who have the power to bind the corporation. - county agent See juvenile officer under officer (1). - del credere agent (del kred-ə-ray or kray-də-ray) (1822) An agent who guarantees the solvency of the third party with whom the agent makes a contract for the principal. • A del credere agent receives possession of the principal's goods for purposes of sale and guarantees that anyone to whom the agent sells the goods on credit will pay promptly for them. For this guaranty, the agent receives a higher commission for sales. The promise of such an agent is almost universally held not to be within the statute of frauds. — Also termed del credere factor. - diplomatic agent (18c) A national representative in one of four categories: (1) ambassadors, (2) envoys and ministers plenipotentiary, (3) ministers resident accredited to the sovereign, or (4) chargés d'affaires accredited to the minister of foreign affairs. - double agent (1935) 1. A spy who finds out an enemy's secrets for his or her principal but who also gives secrets to the enemy. 2. See dual agent (2). - dual agent (1881) 1. See coagent. 2. An agent who represents both parties in a single transaction, esp. a buyer and a seller. — Also termed (in sense 2) double agent. - emigrant agent (1874) One engaged in the business of hiring laborers for work outside the country or state. - enrolled agent See enrolled agent. - escrow agent See escrow agent. - estate agent See real-estate agent. - fiscal agent (18c) A bank or other financial institution that collects and disburses money and services as a depository of private and public funds on another's behalf. - foreign agent (1938) Someone who registers with the federal government as a lobbyist representing the interests of a foreign country or corporation. - forwarding agent (1837) 1. freight forwarder. 2. A freight-forwarder who assembles less-than-carload shipments (small shipments) into carload shipments, thus taking advantage of lower freight rates. - general agent (17c) 1. An agent authorized to transact all the principal's business of a particular kind or in a particular place. • Among the common types of general agents are factors, brokers, and partners. Cf. special agent (1). 2. Insurance. An agent with the general power of making insurance contracts on behalf of an insurer. - government agent (1805) 1. An employee or representative of a governmental body. 2. A law-enforcement official, such as a police officer or an FBI agent. 3. An informant, esp. an inmate, used by law enforcement to obtain incriminating statements from another inmate. - gratuitous agent (1822) An agent who acts without a right to compensation. - high-managerial agent (1957) 1. An agent of a corporation or other business who has authority to formulate corporate policy or supervise employees. — Also termed superior agent. 2. See superior agent (1). - implied agent See apparent agent. - independent agent (17c) An agent who exercises personal judgment and is subject to the principal only for the results of the work performed. Cf. nonservant agent. - innocent agent (1805) Criminal law. A person whose action on behalf of a principal is unlawful but does not merit prosecution because the agent had no knowledge of the principal's illegal purpose; a person who lacks the mens rea for an offense but who is tricked or coerced by the principal into committing a crime. • Although the agent's conduct was unlawful, the agent might not be prosecuted if the agent had no knowledge of the principal's illegal purpose. The principal is legally accountable for the innocent agent's actions. See Model Penal Code § 2.06(2)(a). - insurance agent (1866) Someone authorized by an insurer to sell its policies; specif., an insurer's representative who solicits or procures insurance business, including the continuance, renewal, and revival of policies. — Also termed producer; (in property insurance) recording agent; record agent. - jural agent See jural agent. © 2024 Thomson Reuters. No claim to original U.S. Government Works. 2

AGENT, Black's Law Dictionary (12th ed. 2024) Black's Law Dictionary (12th ed. 2024), agent AGENT Bryan A. Garner, Editor in Chief Preface to the Twelfth Edition | Guide to the Dictionary | Legal Maxims | Bibliography of Books Cited agent (15c) 1. Something that produces an effect <an intervening agent>. See cause (1); electronic agent. 2. Someone who is authorized to act for or in place of another; a representative <a professional athlete's agent>. — Also termed commissionaire. See agency. Cf. principal, n.(1); employee. “Generally speaking, anyone can be an agent who is in fact capable of performing the functions involved. The agent normally binds not himself but his principal by the contracts he makes; it is therefore not essential that he be legally capable to contract (although his duties and liabilities to his principal might be affected by his status). Thus an infant or a lunatic may be an agent, though doubtless the court would disregard either's attempt to act if he were so young or so hopelessly devoid of reason as to be completely incapable of grasping the function he was attempting to perform.” Floyd R. Mechem, Outlines of the Law of Agency 8–9 (Philip Mechem ed., 4th ed. 1952). “The etymology of the word agent or agency tells us much. The words are derived from the Latin verb, ago, agere; the noun agens, agentis. The word agent denotes one who acts, a doer, force or power that accomplishes things.” Harold Gill Reuschlein & William A. Gregory, The Law of Agency and Partnership § 1, at 2–3 (2d ed. 1990). - agent not recognized (2002) Patents. A patent applicant's appointed agent who is not registered to practice before the U.S. Patent and Trademark Office. • A power of attorney appointing an unregistered agent is void. See patent agent. - agent of necessity (1857) An agent that the law empowers to act for the benefit of another in an emergency. — Also termed agent by necessity. - apparent agent (1823) Someone who reasonably appears to have authority to act for another, regardless of whether actual authority has been conferred. — Also termed ostensible agent; implied agent. - associate agent (1993) Patents. An agent who is registered to practice before the U.S. Patent and Trademark Office, has been appointed by a primary agent, and is authorized to prosecute a patent application through the filing of a power of attorney. • An associate agent is often used by outside counsel to assist in-house counsel. See patent agent. - bail-enforcement agent See bounty hunter. - bargaining agent (1935) A labor union in its capacity of representing employees in collective bargaining. - broker-agent See broker. - business agent See business agent. - case agent See case agent. - clearing agent (1937) Securities. A person or company acting as an intermediary in a securities transaction or providing facilities for comparing data regarding securities transactions. • The term includes a custodian of securities in connection with the central handling of securities. Securities Exchange Act § 3(a)(23)(A) (15 USCA § 78c(a)(23)(A)). — Also termed clearing agency. - closing agent (1922) An agent who represents the purchaser or buyer in the negotiation and closing of a real-property transaction by handling financial calculations and transfers of documents. — Also termed settlement agent. See also settlement attorney under attorney. - coagent (16c) Someone who shares with another agent the authority to act for the principal. • A coagent may be appointed by the principal or by another agent who has been authorized to make the appointment. — Also termed dual agent. Cf. common agent. - commercial agent (18c) 1. broker. 2. A consular officer responsible for the commercial interests of his or her country at a foreign port. 3. See mercantile agent. 4. See commission agent. © 2024 Thomson Reuters. No claim to original U.S. Government Works. 1

38 Chapter 2 Intelligent Agents A B Figure2.2 Avacuum-cleanerworldwithjusttwolocations. Eachlocationcanbecleanor dirty,andtheagentcanmoveleftorrightandcancleanthesquarethatitoccupies.Different versions of the vacuum world allow for different rules about what the agent can perceive, whetheritsactionsalwayssucceed,andsoon. Perceptsequence Action [A,Clean] Right [A,Dirty] Suck [B,Clean] Left [B,Dirty] Suck [A,Clean],[A,Clean] Right [A,Clean],[A,Dirty] Suck . . . . . . [A,Clean],[A,Clean],[A,Clean] Right [A,Clean],[A,Clean],[A,Dirty] Suck . . . . . . Figure2.3 Partialtabulationofasimpleagentfunctionforthevacuum-cleanerworldshown in Figure2.2.Theagentcleansthecurrentsquareifitisdirty,otherwiseitmovestotheother square. Notethatthetableisofunboundedsizeunlessthereisarestrictiononthelengthof possibleperceptsequences. Before closing this section, weshould emphasize that thenotion ofanagent ismeant to be a tool for analyzing systems, not an absolute characterization that divides the world into agents and non-agents. One could view a hand-held calculator as an agent that chooses the action of displaying “4” when given the percept sequence “2 + 2 =,” but such an analysis wouldhardlyaidourunderstanding ofthecalculator. Inasense, allareasofengineering can be seen as designing artifacts that interact with the world; AI operates at (what the authors consider to be) the most interesting end of the spectrum, where the artifacts have significant computational resources andthetaskenvironment requiresnontrivial decision making.

62 Chapter 2 Intelligent Agents Asnoted in Chapter 1, the development of utility theory as a basis for rational behavior goes back hundreds of years. In AI, early research eschewed utilities in favor of goals, with some exceptions (Feldman and Sproull, 1977). The resurgence of interest in probabilistic methods in the 1980s led to the acceptance of maximization of expected utility as the most general framework for decision making (Horvitz etal., 1988). The text by Pearl(1988) was thefirstin AItocoverprobabilityandutilitytheoryindepth;itsexpositionofpracticalmethods for reasoning and decision making under uncertainty was probably the single biggest factor in the rapid shift towards utility-based agents in the 1990s (see Chapter 16). The formalization of reinforcement learning within adecision-theoretic framework also contributed tothisshift(Sutton, 1988). Somewhatremarkably, almostall AIresearch until veryrecently hasassumedthattheperformance measurecanbeexactlyandcorrectlyspecifiedintheform ofautilityfunctionorrewardfunction(Hadfield-Menell etal.,2017a;Russell,2019). Thegeneral design forlearning agents portrayed in Figure2.15isclassic inthemachine learning literature (Buchanan et al., 1978; Mitchell, 1997). Examples of the design, as embodiedinprograms,gobackatleastasfaras Arthur Samuel’s(1959,1967)learningprogram forplayingcheckers. Learningagentsarediscussed indepthin Chapters19–22. Someearly papers on agent-based approaches are collected by Huhns and Singh (1998) and Wooldridgeand Rao(1999). Textsonmultiagentsystemsprovideagoodintroduction to many aspects of agent design (Weiss, 2000a; Wooldridge, 2009). Several conference series devoted to agents began in the 1990s, including the International Workshop on Agent Theories, Architectures, and Languages (ATAL), the International Conference on Autonomous Agents (AGENTS),and the International Conference on Multi-Agent Systems (ICMAS).In 2002, thesethree mergedtoformthe International Joint Conference on Autonomous Agents and Multi-Agent Systems (AAMAS). From 2000 to 2012 there were annual workshops on Agent-Oriented Software Engineering (AOSE). The journal Autonomous Agents and Multi Agent Systems was founded in 1998. Finally, Dung Beetle Ecology (Hanski and Cambefort, 1991)providesawealthofinterestinginformation onthebehaviorofdungbeetles. You Tube hasinspiring videorecordings oftheiractivities.

Bibliographicaland Historical Notes 61 concept of a controller in control theory is identical to that of an agent in AI. Perhaps sur- Controller prisingly, AI has concentrated for most of its history on isolated components of agents— question-answering systems, theorem-provers, vision systems, and so on—rather than on wholeagents. Thediscussion ofagents inthetextby Genesereth and Nilsson(1987) wasan influentialexception. Thewhole-agentviewisnowwidelyacceptedandisacentralthemein recenttexts(Padghamand Winikoff, 2004;Jones,2007;Pooleand Mackworth,2017). Chapter 1traced the roots ofthe concept of rationality in philosophy and economics. In AI,theconceptwasofperipheralinterestuntilthemid-1980s,whenitbegantosuffusemany discussions aboutthepropertechnical foundations ofthefield. Apaperby Jon Doyle(1983) predicted that rational agent design would come to be seen as the core mission of AI, while otherpopulartopicswouldspinofftoformnewdisciplines. Careful attention to the properties of the environment and their consequences for rational agent design is most apparent in the control theory tradition—for example, classical control systems (Dorf and Bishop, 2004; Kirk, 2004) handle fully observable, deterministic environments; stochastic optimal control (Kumar and Varaiya, 1986; Bertsekas and Shreve, 2007) handles partially observable, stochastic environments; and hybrid control (Henzinger and Sastry, 1998; Cassandras and Lygeros, 2006) deals with environments containing both discrete andcontinuous elements. Thedistinction betweenfully andpartially observable environments is also central in the dynamic programming literature developed in the field of operations research (Puterman,1994), whichwediscuss in Chapter17. Although simple reflex agents were central to behaviorist psychology (see Chapter 1), most AIresearchersviewthemastoosimpletoprovidemuchleverage. (Rosenschein (1985) and Brooks (1986) questioned this assumption; see Chapter 26.) A great deal of work has gone into finding efficient algorithms for keeping track of complex environments (Bar Shalometal.,2001;Chosetetal.,2005;Simon,2006),mostofitintheprobabilistic setting. Goal-based agents are presupposed in everything from Aristotle’s view of practical reasoning to Mc Carthy’s early papers on logical AI. Shakey the Robot (Fikes and Nilsson, 1971; Nilsson, 1984) was the first robotic embodiment of a logical, goal-based agent. A full logical analysis of goal-based agents appeared in Genesereth and Nilsson (1987), and a goal-basedprogrammingmethodologycalledagent-orientedprogrammingwasdevelopedby Shoham (1993). The agent-based approach is now extremely popular in software engineering (Ciancarini and Wooldridge, 2001). It has also infiltrated the area of operating systems, whereautonomiccomputingreferstocomputersystemsandnetworksthatmonitorandcon- Autonomic computing trolthemselves withaperceive–act loopandmachinelearning methods(Kephartand Chess, 2003). Noting that a collection of agent programs designed to work well together in a true multiagentenvironmentnecessarilyexhibitsmodularity—theprogramssharenointernalstate and communicate with each other only through the environment—it is common within the field of multiagent systems to design the agent program of asingle agent as a collection of autonomous sub-agents. In some cases, one can even prove that the resulting system gives thesameoptimalsolutions asamonolithicdesign. The goal-based view of agents also dominates the cognitive psychology tradition in the area of problem solving, beginning with the enormously influential Human Problem Solving(Newelland Simon,1972)andrunningthroughallof Newell’slaterwork(Newell,1990). Goals, further analyzed asdesires (general) andintentions (currently pursued), arecentral to theinfluentialtheoryofagentsdevelopedby Michael Bratman(1987).

60 Chapter 2 Intelligent Agents if the representation of a concept is spread over many memory locations, and each memory location is employed as part of the representation of multiple different concepts, we call Distributed thatadistributedrepresentation. Distributed representations aremorerobust against noise representation and information loss. With a localist representation, the mapping from concept to memory location is arbitrary, and if a transmission error garbles a few bits, we might confuse Truck withtheunrelatedconcept Truce. Butwithadistributedrepresentation, youcanthinkofeach conceptrepresentingapointinmultidimensionalspace,andifyougarbleafewbitsyoumove toanearbypointinthatspace,whichwillhavesimilarmeaning. Summary This chapter has been something of a whirlwind tour of AI, which we have conceived of as thescienceofagentdesign. Themajorpointstorecallareasfollows: • Anagent issomething that perceives and acts in an environment. The agent function foranagentspecifiestheactiontakenbytheagentinresponsetoanyperceptsequence. • The performance measure evaluates the behavior of the agent in an environment. A rational agentactssoastomaximize theexpected valueoftheperformance measure, giventheperceptsequence ithasseensofar. • A task environment specification includes the performance measure, the external environment, the actuators, and the sensors. In designing an agent, the first step must alwaysbetospecifythetaskenvironment asfullyaspossible. • Taskenvironments varyalongseveralsignificantdimensions. Theycanbefullyorpartiallyobservable,single-agentormultiagent,deterministicornondeterministic,episodic orsequential, staticordynamic, discreteorcontinuous, andknownorunknown. • Incaseswheretheperformance measureisunknown orhardtospecify correctly, there isasignificantriskoftheagentoptimizingthewrongobjective. Insuchcasestheagent designshouldreflectuncertainty aboutthetrueobjective. • The agent program implements the agent function. There exists a variety of basic agent program designs reflecting thekindofinformation madeexplicit and usedinthe decision process. The designs vary in efficiency, compactness, and flexibility. The appropriate designoftheagentprogram depends onthenatureoftheenvironment. • Simplereflexagentsresponddirectlytopercepts,whereasmodel-basedreflexagents maintain internal state to track aspects of the world that are not evident in the current percept. Goal-based agents act to achieve their goals, and utility-based agents try to maximizetheirownexpected “happiness.” • Allagentscanimprovetheirperformance throughlearning. Bibliographical and Historical Notes The central role of action in intelligence—the notion of practical reasoning—goes back at least as far as Aristotle’s Nicomachean Ethics. Practical reasoning was also the subject of Mc Carthy’sinfluential paper“Programswith Common Sense”(1958). Thefieldsofrobotics andcontrol theory are, bytheir verynature, concerned principally withphysical agents. The

Section2.4 The Structureof Agents 59 In an atomic representation each state of the world is indivisible—it has no internal Atomic representation structure. Consider thetask offindingadriving route from oneendofacountry totheother viasomesequence ofcities(weaddress thisproblem in Figure3.1onpage64). Forthepurposesofsolvingthisproblem,itmaysufficetoreducethestateoftheworldtojustthenameof thecity wearein—asingle atom ofknowledge, a“black box” whoseonly discernible propertyisthatofbeingidenticaltoordifferentfromanotherblackbox. Thestandardalgorithms underlying search and game-playing (Chapters 3–5), hidden Markov models (Chapter 14), and Markovdecision processes (Chapter17)allworkwithatomicrepresentations. Afactoredrepresentationsplitsupeachstateintoafixedsetofvariablesorattributes, Factored representation each of which can have a value. Consider a higher-fidelity description for the same driving Variable problem, where we need to be concerned with more than just atomic location in one city or Attribute another; we might need to pay attention to how much gas is in the tank, our current GPS Value coordinates, whether or not the oil warning light is working, how much money we have for tolls,whatstationisontheradio,andsoon. Whiletwodifferentatomicstateshavenothingin common—they are just different black boxes—two different factored states can share some attributes (suchasbeingatsomeparticular GPSlocation) andnotothers(suchashaving lots ofgasorhaving nogas);thismakesitmucheasiertoworkouthowtoturnonestateintoanother. Manyimportantareasof AIarebasedonfactoredrepresentations, includingconstraint satisfaction algorithms (Chapter 6), propositional logic (Chapter 7), planning (Chapter 11), Bayesiannetworks(Chapters12–16), andvariousmachinelearning algorithms. For many purposes, we need to understand the world as having things in it that are related to each other, not just variables withvalues. Forexample, wemightnotice that alarge truck ahead of us is reversing into the driveway of a dairy farm, but a loose cow is blocking the truck’s path. A factored representation is unlikely to be pre-equipped with the attribute Truck Ahead Backing Into Dairy Farm Driveway Blocked By Loose Cow with value true or false. Instead, we would need a structured representation, in which objects such as cows Structured representation and trucks and their various and varying relationships can be described explicitly (see Figure 2.16(c)). Structured representations underlie relational databases and first-order logic (Chapters8,9,and10),first-orderprobability models(Chapter15),andmuchofnaturallanguage understanding (Chapters 23 and24). Infact, muchofwhathumans express innatural language concerns objectsandtheirrelationships. As we mentioned earlier, the axis along which atomic, factored, and structured representations lieistheaxis ofincreasing expressiveness. Roughly speaking, amoreexpressive Expressiveness representation cancapture,atleastasconcisely, everythingalessexpressiveonecancapture, plussomemore. Often,themoreexpressivelanguageismuchmoreconcise;forexample,the rules of chess can be written in a page or two of a structured-representation language such as first-order logic but require thousands of pages when written in a factored-representation language such as propositional logic and around 1038 pages when written in an atomic languagesuchasthatoffinite-stateautomata. Ontheotherhand,reasoningandlearningbecome more complex as the expressive power of the representation increases. To gain the benefits ofexpressive representations whileavoiding their drawbacks, intelligent systems forthereal worldmayneedtooperateatallpointsalongtheaxissimultaneously. Anotheraxisforrepresentation involvesthemappingofconceptstolocationsinphysical memory, whether in a computer or in a brain. If there is a one-to-one mapping between concepts and memory locations, we call that a localist representation. On the other hand, Localist representation

58 Chapter 2 Intelligent Agents More generally, human choices can provide information about human preferences. For example, suppose the taxi does not know that people generally don’t like loud noises, and settles on the idea of blowing its horn continuously as a way of ensuring that pedestrians knowit’scoming. Theconsequent humanbehavior—covering ears,usingbadlanguage, and possibly cutting the wires to the horn—would provide evidence to the agent with which to updateitsutilityfunction. Thisissueisdiscussed furtherin Chapter22. In summary, agents have a variety of components, and those components can be represented in many ways within the agent program, so there appears to be great variety among learning methods. Thereis, however, asingle unifying theme. Learning inintelligent agents canbesummarized asaprocess ofmodification ofeachcomponent oftheagent tobring the components into closer agreement withthe available feedback information, thereby improvingtheoverallperformance oftheagent. 2.4.7 How the components of agent programs work Wehavedescribedagentprograms(inveryhigh-levelterms)asconsistingofvariouscomponents,whosefunctionitistoanswerquestionssuchas: “Whatistheworldlikenow?” “What action should I do now?” “What do my actions do?” The next question for a student of AI is, “How on Earth do these components work?” It takes about a thousand pages to begin to answer that question properly, buthere wewantto draw thereader’s attention to somebasic distinctions among thevarious waysthatthecomponents canrepresent theenvironment that theagentinhabits. Roughlyspeaking,wecanplacetherepresentations alonganaxisofincreasingcomplexityandexpressive power—atomic, factored, andstructured. Toillustrate theseideas, ithelps to consider a particular agent component, such as the one that deals with “What my actions do.” Thiscomponent describes thechanges thatmightoccur intheenvironment astheresult of taking an action, and Figure 2.16 provides schematic depictions of how those transitions mightberepresented. B C B C (a) Atomic (b) Factored (c) Structured Figure 2.16 Three ways to represent states and the transitions between them. (a) Atomic representation:astate(suchas Bor C)isablackboxwithnointernalstructure;(b)Factored representation: a state consists of a vectorof attribute values; valuescan be Boolean, realvalued, or one of a fixed set of symbols. (c) Structured representation: a state includes objects,eachofwhichmayhaveattributesofitsownaswellasrelationshipstootherobjects.

Section2.4 The Structureof Agents 57 tobetheentire agent: ittakes inpercepts and decides onactions. Thelearning element uses feedback from the critic on how the agent is doing and determines how the performance Critic elementshouldbemodifiedtodobetterinthefuture. Thedesignofthelearning elementdependsverymuchonthedesignoftheperformance element. When trying to design an agent that learns a certain capability, the first question is not“Howam Igoingtogetittolearnthis?” but“Whatkindofperformanceelementwillmy agent use to do this once it has learned how?” Given a design for the performance element, learning mechanismscanbeconstructed toimproveeverypartoftheagent. The critic tells the learning element how well the agent is doing with respect to a fixed performance standard. The critic is necessary because the percepts themselves provide no indication of the agent’s success. For example, a chess program could receive a percept indicating that it has checkmated its opponent, but it needs a performance standard to know thatthisisagoodthing;theperceptitselfdoesnotsayso. Itisimportantthattheperformance standard be fixed. Conceptually, one should think of it as being outside the agent altogether becausetheagentmustnotmodifyittofititsownbehavior. The last component of the learning agent is the problem generator. It is responsible Problem generator for suggesting actions that willlead to new and informative experiences. If the performance element had its way, it would keep doing the actions that are best, given what it knows, but iftheagent iswilling toexplore alittle anddosomeperhaps suboptimal actions intheshort run,itmightdiscovermuchbetteractions forthelongrun. Theproblemgenerator’s jobisto suggesttheseexploratoryactions. Thisiswhatscientistsdowhentheycarryoutexperiments. Galileodidnotthinkthatdroppingrocksfromthetopofatowerin Pisawasvaluableinitself. Hewasnot trying to break the rocks orto modify the brains of unfortunate pedestrians. His aimwastomodifyhisownbrainbyidentifying abettertheoryofthemotionofobjects. The learning element can make changes to any of the “knowledge” components shown intheagentdiagrams(Figures2.9,2.11,2.13,and2.14). Thesimplestcasesinvolvelearning directly from the percept sequence. Observation of pairs of successive states of the environment can allow the agent to learn “What my actions do” and “How the world evolves” in response to its actions. For example, if the automated taxi exerts a certain braking pressure when driving on a wet road, then it will soon find out how much deceleration is actually achieved, and whether it skids off the road. The problem generator might identify certain parts of the model that are in need of improvement and suggest experiments, such as trying outthebrakesondifferent roadsurfaces underdifferentconditions. Improving the model components of a model-based agent so that they conform better with reality is almost always a good idea, regardless of the external performance standard. (In some cases, it is better from a computational point of view to have a simple but slightly inaccurate model rather than a perfect but fiendishly complex model.) Information from the externalstandard isneededwhentryingtolearnareflexcomponent orautilityfunction. For example, suppose the taxi-driving agent receives no tips from passengers who have been thoroughly shaken up during the trip. The external performance standard must inform the agent that the loss of tips is a negative contribution to its overall performance; then the agent might be able to learn that violent maneuvers do not contribute to its own utility. In a sense, the performance standard distinguishes part of the incoming percept as a reward Reward (orpenalty)thatprovides directfeedback onthequality oftheagent’s behavior. Hard-wired Penalty performance standards suchaspainandhungerinanimalscanbeunderstood inthisway.

56 Chapter 2 Intelligent Agents Performance standard Agent Environment Critic Sensors feedback changes Learning Performance element element knowledge learning goals Problem generator Actuators Figure2.15 Agenerallearningagent. The“performanceelement”boxrepresentswhatwe havepreviouslyconsideredtobethewholeagentprogram.Now,the“learningelement”box getstomodifythatprogramtoimproveitsperformance. unachievable inpracticebecauseofcomputational complexity,aswenotedin Chapter1. We alsonotethatnotallutility-based agentsaremodel-based; wewillseein Chapters22and26 Model-freeagent that a model-free agent can learn what action is best in a particular situation without ever learning exactlyhowthatactionchangestheenvironment. Finally, all of this assumes that the designer can specify the utility function correctly; Chapters17,18,and22consider theissueofunknownutilityfunctions inmoredepth. 2.4.6 Learning agents We have described agent programs with various methods for selecting actions. We have not, so far, explained how the agent programs come into being. In his famous early paper, Turing (1950) considers the idea of actually programming his intelligent machines by hand. Heestimateshowmuchworkthismighttakeandconcludes,“Somemoreexpeditiousmethod seems desirable.” The method he proposes is to build learning machines and then to teach them. In many areas of AI, this is now the preferred method for creating state-of-the-art systems. Any type of agent (model-based, goal-based, utility-based, etc.) can be built as a learning agent(ornot). Learninghasanotheradvantage, aswenotedearlier: itallowstheagenttooperateininitiallyunknownenvironments andtobecomemorecompetentthanitsinitialknowledgealone mightallow. Inthissection,webrieflyintroducethemainideasoflearningagents. Throughout the book, we comment on opportunities and methods for learning in particular kinds of agents. Chapters19–22gointomuchmoredepthonthelearning algorithms themselves. A learning agent can be divided into four conceptual components, as shown in Fig Learningelement ure 2.15. The most important distinction is between the learning element, which is re Performance sponsibleformakingimprovements,andtheperformanceelement,whichisresponsiblefor element selecting external actions. The performance element is what we have previously considered

48 Chapter 2 Intelligent Agents function TABLE-DRIVEN-AGENT(percept)returnsanaction persistent: percepts,asequence,initiallyempty table,atableofactions,indexedbyperceptsequences,initiallyfullyspecified appendpercepttotheendofpercepts action←LOOKUP(percepts,table) returnaction Figure2.7 The TABLE-DRIVEN-AGENT programisinvokedforeach newperceptandreturnsanactioneachtime. Itretainsthecompleteperceptsequenceinmemory. 2.4.1 Agent programs The agent programs that we design in this book all have the same skeleton: they take the current percept as input from the sensors and return an action to the actuators.5 Notice the differencebetweentheagentprogram,whichtakesthecurrentperceptasinput,andtheagent function, which maydepend on theentire percept history. The agent program has no choice but totake just the current percept asinput because nothing moreisavailable from the environment; iftheagent’s actions need todepend onthe entire percept sequence, theagent will havetorememberthepercepts. We describe the agent programs in the simple pseudocode language that is defined in Appendix B. (The online code repository contains implementations in real programming languages.) Forexample, Figure 2.7shows arather trivial agent program thatkeeps track of the percept sequence and then uses it to index into a table of actions to decide what to do. The table—an example of which is given for the vacuum world in Figure 2.3—represents explicitly the agent function that the agent program embodies. To build a rational agent in thisway,weasdesignersmustconstructatablethatcontainstheappropriateactionforevery possible perceptsequence. Itisinstructivetoconsiderwhythetable-drivenapproachtoagentconstructionisdoomed tofailure. Let P bethesetofpossibleperceptsandlet T bethelifetimeoftheagent(thetotal numberofperceptsitwillreceive). Thelookuptablewillcontain∑T |P|t entries. Consider t=1 the automated taxi: the visual input from a single camera (eight cameras is typical) comes in at the rate of roughly 70 megabytes per second (30 frames per second, 1080×720 pixels with24bitsofcolorinformation). Thisgivesalookuptablewithover10600,000,000,000 entries foranhour’s driving. Eventhelookup table forchess—a tiny, well-behaved fragment ofthe real world—has (it turns out) at least 10150 entries. In comparison, the number of atoms in theobservable universe islessthan1080. Thedaunting sizeofthesetables meansthat(a)no physical agent in this universe will have the space to store the table; (b) the designer would not have time to create the table; and (c) no agent could ever learn all the right table entries fromitsexperience. ◮ Despite all this, TABLE-DRIVEN-AGENT does do what we want, assuming the table is filledincorrectly: itimplementsthedesiredagentfunction. 5 Thereareotherchoices fortheagent program skeleton; forexample, wecouldhavetheagent programsbe coroutinesthatrunasynchronouslywiththeenvironment. Eachsuchcoroutinehasaninputandoutputportand consistsofaloopthatreadstheinputportforperceptsandwritesactionstotheoutputport.

CHAPTER INTELLIGENT AGENTS Inwhichwediscussthenatureofagents,perfectorotherwise,thediversityofenvironments, andtheresultingmenagerie ofagenttypes. Chapter 1 identified the concept of rational agents as central to our approach to artificial intelligence. Inthischapter,wemakethisnotionmoreconcrete. Wewillseethattheconcept ofrationality canbeappliedtoawidevarietyofagentsoperating inanyimaginable environment. Ourplan inthisbook istousethis concept todevelop asmallset ofdesign principles forbuilding successful agents—systems thatcanreasonably becalledintelligent. Webegin by examining agents, environments, and the coupling between them. The observation that some agents behave better than others leads naturally to the idea of a rational agent—one that behaves as well as possible. How well an agent can behave depends on the natureoftheenvironment;someenvironmentsaremoredifficultthanothers. Wegiveacrude categorization ofenvironments andshowhowproperties ofanenvironment influencethedesignofsuitable agentsforthatenvironment. Wedescribeanumberofbasic“skeleton” agent designs, whichwefleshoutintherestofthebook. 2.1 Agents and Environments Environment Anagentisanything thatcanbeviewedasperceiving itsenvironmentthrough sensorsand Sensor actinguponthatenvironmentthroughactuators. Thissimpleideaisillustratedin Figure2.1. Actuator A human agent has eyes, ears, and other organs for sensors and hands, legs, vocal tract, and so on for actuators. A robotic agent might have cameras and infrared range finders for sensors and various motors for actuators. A software agent receives file contents, network packets, andhumaninput(keyboard/mouse/touchscreen/voice) assensory inputsandactson the environment by writing files, sending network packets, and displaying information or generating sounds. Theenvironment could beeverything—the entire universe! Inpractice it isjustthatpartoftheuniversewhosestatewecareaboutwhendesigningthisagent—thepart thataffectswhattheagentperceivesandthatisaffectedbytheagent’s actions. Percept We use the term percept to refer to the content an agent’s sensors are perceiving. An Percept sequenc◮e agent’sperceptsequenceisthecompletehistoryofeverything theagenthaseverperceived. Ingeneral, anagent’s choice ofaction atanygiven instant candepend onitsbuilt-in knowledge and on the entire percept sequence observed to date, but not on anything it hasn’t perceived. By specifying the agent’s choice of action for every possible percept sequence, we have said more or less everything there is to say about the agent. Mathematically speak Agentfunction ing, wesay that an agent’s behavior is described by the agent function that maps any given perceptsequence toanaction.

Section2.4 The Structureof Agents 55 Agent Environment Sensors State What the world How the world evolves is like now What it will be like What my actions do if I do action A How happy I will be Utility in such a state What action I should do now Actuators Figure2.14 Amodel-based,utility-basedagent. Itusesamodeloftheworld,alongwitha utilityfunctionthatmeasuresitspreferencesamongstatesoftheworld. Thenitchoosesthe actionthatleadstothebestexpectedutility,whereexpectedutilityiscomputedbyaveraging overallpossibleoutcomestates,weightedbytheprobabilityoftheoutcome. aim for, none of which can be achieved with certainty, utility provides a way in which the likelihood ofsuccesscanbeweighedagainsttheimportance ofthegoals. Partial observability and nondeterminism areubiquitous inthereal world, and so, therefore, is decision making under uncertainty. Technically speaking, a rational utility-based agent chooses the action that maximizes the expected utility of the action outcomes—that Expected utility is, the utility the agent expects to derive, on average, given the probabilities and utilities of each outcome. (Appendix A defines expectation more precisely.) In Chapter 16, we show thatanyrational agentmustbehave asifitpossesses autilityfunction whoseexpected value ittriestomaximize. Anagentthatpossessesanexplicitutilityfunctioncanmakerationaldecisionswithageneral-purpose algorithmthatdoesnotdependonthespecificutilityfunction being maximized. Inthis way, the “global” definition ofrationality—designating asrational those agent functions that have the highest performance—is turned into a “local” constraint onrational-agent designs thatcanbeexpressed inasimpleprogram. The utility-based agent structure appears in Figure 2.14. Utility-based agent programs appearin Chapters16and17,wherewedesigndecision-making agentsthatmusthandle the uncertaintyinherentinnondeterministicorpartiallyobservableenvironments. Decisionmakinginmultiagentenvironmentsisalsostudiedintheframeworkofutilitytheory,asexplained in Chapter18. Atthis point, the reader maybe wondering, “Is it that simple? Wejust build agents that maximize expected utility, and we’re done?” It’s true that such agents would be intelligent, but it’s not simple. A utility-based agent has to model and keep track of its environment, tasks that have involved a great deal of research on perception, representation, reasoning, and learning. The results of this research fill many of the chapters of this book. Choosing theutility-maximizing courseofactionisalsoadifficulttask,requiringingeniousalgorithms that fill several more chapters. Even with these algorithms, perfect rationality is usually

54 Chapter 2 Intelligent Agents Agent Environment Sensors State What the world How the world evolves is like now What it will be like What my actions do if I do action A What action I Goals should do now Actuators Figure 2.13 A model-based,goal-basedagent. It keepstrack of the world state as well as asetofgoalsitistryingtoachieve,andchoosesanactionthatwill(eventually)leadtothe achievementofitsgoals. simply by specifying that destination as the goal. The reflex agent’s rules for when to turn and when to go straight will work only for a single destination; they must all be replaced to gosomewherenew. 2.4.5 Utility-based agents Goals alone are not enough to generate high-quality behavior in most environments. For example, many action sequences will get the taxi to its destination (thereby achieving the goal), butsomearequicker, safer,morereliable, orcheaper thanothers. Goalsjustprovide a crudebinarydistinctionbetween“happy”and“unhappy”states. Amoregeneralperformance measureshouldallowacomparison ofdifferentworldstatesaccordingtoexactlyhowhappy theywouldmaketheagent. Because“happy” doesnotsoundveryscientific, economists and Utility computerscientists usethetermutilityinstead.7 Wehave already seenthat aperformance measure assigns ascoretoanygivensequence of environment states, so it can easily distinguish between more and less desirable ways of Utilityfunction getting to the taxi’s destination. An agent’s utility function is essentially an internalization of the performance measure. Provided that the internal utility function and the external performancemeasureareinagreement,anagentthatchoosesactionstomaximizeitsutilitywill berationalaccording totheexternalperformance measure. Letusemphasizeagainthatthisisnottheonlywaytoberational—wehavealreadyseen a rational agent program for the vacuum world (Figure 2.8) that has no idea what its utility function is—but, like goal-based agents, autility-based agent has many advantages in terms of flexibility and learning. Furthermore, in two kinds of cases, goals are inadequate but a utility-based agent can still make rational decisions. First, when there are conflicting goals, only some of which can be achieved (for example, speed and safety), the utility function specifies the appropriate tradeoff. Second, when there are several goals that the agent can 7 Theword“utility”hererefersto“thequalityofbeinguseful,”nottotheelectriccompanyorwaterworks.

Section2.4 The Structureof Agents 53 function MODEL-BASED-REFLEX-AGENT(percept)returnsanaction persistent: state,theagent’scurrentconceptionoftheworldstate transition model,adescriptionofhowthenextstatedependson thecurrentstateandaction sensor model,adescriptionofhowthecurrentworldstateisreflected intheagent’spercepts rules,asetofcondition–actionrules action,themostrecentaction,initiallynone state←UPDATE-STATE(state,action,percept,transition model,sensor model) rule←RULE-MATCH(state,rules) action←rule.ACTION returnaction Figure 2.12 A model-based reflex agent. It keeps track of the current state of the world, usinganinternalmodel.Itthenchoosesanactioninthesamewayasthereflexagent. maynotbeabletoseearoundthelargetruckthathasstoppedinfrontofitandcanonlyguess about what may be causing the hold-up. Thus, uncertainty about the current state may be unavoidable, buttheagentstillhastomakeadecision. 2.4.4 Goal-based agents Knowingsomethingaboutthecurrentstateoftheenvironmentisnotalwaysenoughtodecide what to do. For example, at a road junction, the taxi can turn left, turn right, or go straight on. The correct decision depends on where the taxi is trying to get to. In other words, as well as a current state description, the agent needs some sort of goal information that Goal describes situations that are desirable—for example, being at a particular destination. The agent program can combine this with the model (the same information as was used in the model-based reflex agent) to choose actions that achieve the goal. Figure 2.13 shows the goal-based agent’sstructure. Sometimes goal-based action selection is straightforward—for example, when goal satisfaction results immediately from a single action. Sometimes it will be more tricky—for example,whentheagenthastoconsider longsequences oftwistsandturnsinordertofinda waytoachievethegoal. Search(Chapters3to5)andplanning(Chapter11)arethesubfields of AIdevoted tofindingactionsequences thatachievetheagent’sgoals. Notice that decision making of this kind is fundamentally different from the condition– actionrulesdescribed earlier,inthatitinvolvesconsideration ofthefuture—both“Whatwill happen if Idosuch-and-such?” and“Willthatmakemehappy?” Inthereflexagentdesigns, this information is not explicitly represented, because the built-in rules map directly from percepts to actions. The reflex agent brakes when it sees brake lights, period. It has no idea why. A goal-based agent brakes whenit sees brake lights because that’s the only action that itpredicts willachieveitsgoalofnothittingothercars. Although the goal-based agent appears less efficient, it is more flexible because the knowledge that supports its decisions is represented explicitly and can be modified. For example,agoal-basedagent’sbehaviorcaneasilybechangedtogotoadifferentdestination,

52 Chapter 2 Intelligent Agents Agent Environment Sensors State How the world evolves What the world is like now What my actions do What action I Condition-action rules should do now Actuators Figure2.11 Amodel-basedreflexagent. Updatingthisinternalstateinformationastimegoesbyrequirestwokindsofknowledge tobeencodedintheagentprograminsomeform. First,weneedsomeinformationabouthow the world changes over time, which can be divided roughly into twoparts: the effects of the agent’sactionsandhowtheworldevolvesindependentlyoftheagent. Forexample,whenthe agent turns the steering wheelclockwise, the car turns to theright, and when it’s raining the car’s cameras can get wet. This knowledge about “how the world works”—whether implemented in simple Boolean circuits or in complete scientific theories—is called a transition Transitionmodel modeloftheworld. Second, we need some information about how the state of the world is reflected in the agent’s percepts. For example, when the car in front initiates braking, one or more illuminated red regions appear in the forward-facing camera image, and, when the camera gets wet, droplet-shaped objects appear in the image partially obscuring the road. This kind of Sensormodel knowledgeiscalledasensormodel. Together, thetransition modelandsensormodelallowanagenttokeeptrackofthestate of the world—to the extent possible given the limitations of the agent’s sensors. An agent Model-basedagent thatusessuchmodelsiscalledamodel-basedagent. Figure2.11givesthestructure ofthemodel-based reflexagentwithinternal state,showing how the current percept is combined with the old internal state to generate the updated descriptionofthecurrentstate,basedontheagent’smodelofhowtheworldworks. Theagent programisshownin Figure2.12. Theinterestingpartisthefunction UPDATE-STATE,which is responsible for creating the new internal state description. Thedetails of how models and states are represented vary widely depending on the type of environment and the particular technology usedintheagentdesign. Regardlessofthekindofrepresentation used,itisseldompossiblefortheagenttodeterminethecurrent stateofapartially observable environment exactly. Instead, theboxlabeled “whattheworldislikenow”(Figure2.11)represents theagent’s“bestguess”(orsometimes best guesses, if the agent entertains multiple possibilities). For example, an automated taxi

Section2.4 The Structureof Agents 51 function SIMPLE-REFLEX-AGENT(percept)returnsanaction persistent: rules,asetofcondition–actionrules state←INTERPRET-INPUT(percept) rule←RULE-MATCH(state,rules) action←rule.ACTION returnaction Figure2.10 Asimplereflexagent. Itactsaccordingtoarulewhoseconditionmatchesthe currentstate,asdefinedbythepercept. Even a little bit of unobservability can cause serious trouble. For example, the braking rule given earlier assumes that the condition car-in-front-is-braking can be determined from the current percept—a single frame of video. This works if the car in front has a centrally mounted (and hence uniquely identifiable) brake light. Unfortunately, older models have different configurations of taillights, brake lights, and turn-signal lights, and it is not always possible to tell from a single image whether the car is braking or simply has its taillights on. A simple reflex agent driving behind such a car would either brake continuously and unnecessarily, or,worse,neverbrakeatall. Wecan seeasimilar problem arising inthe vacuum world. Suppose that asimple reflex vacuum agent is deprived of its location sensor and has only a dirt sensor. Such an agent has just two possible percepts: [Dirty] and [Clean]. It can Suck in response to [Dirty]; what shoulditdoinresponseto[Clean]? Moving Leftfails(forever)ifithappenstostartinsquare A, and moving Right fails (forever) ifithappens to start insquare B. Infinite loops are often unavoidable forsimplereflexagentsoperating inpartially observable environments. Escape from infinite loops ispossible ifthe agent canrandomize itsactions. Forexam- Randomization ple, if the vacuum agent perceives [Clean], it might flip a coin to choose between Right and Left. It is easy to show that the agent will reach the other square in an average of twosteps. Then, if that square is dirty, the agent will clean it and the task will be complete. Hence, a randomized simplereflexagentmightoutperform adeterministic simplereflexagent. Wementionedin Section2.3thatrandomizedbehavioroftherightkindcanberationalin some multiagent environments. In single-agent environments, randomization is usually not rational. It is a useful trick that helps a simple reflex agent in some situations, but in most caseswecandomuchbetterwithmoresophisticated deterministic agents. 2.4.3 Model-based reflex agents The most effective way to handle partial observability is for the agent to keep track of the part of the world it can’t see now. That is, the agent should maintain some sort of internal statethatdepends onthepercepthistoryandtherebyreflectsatleastsomeoftheunobserved Internal state aspects ofthecurrentstate. Forthebraking problem, theinternal stateisnottooextensive— just the previous frame from the camera, allowing the agent to detect when twored lights at theedgeofthevehicle goonoroffsimultaneously. Forotherdriving taskssuchaschanging lanes,theagentneedstokeeptrackofwheretheothercarsareifitcan’tseethemallatonce. Andforanydrivingtobepossibleatall,theagentneedstokeeptrackofwhereitskeysare.

Section2.2 Good Behavior:The Conceptof Rationality 39 2.2 Good Behavior: The Concept of Rationality A rational agent is one that does the right thing. Obviously, doing the right thing is better Rationalagent thandoingthewrongthing,butwhatdoesitmeantodotherightthing? 2.2.1 Performance measures Moral philosophy has developed several different notions of the “right thing,” but AI has generallystucktoonenotioncalledconsequentialism: weevaluateanagent’sbehaviorbyits Consequentialism consequences. Whenanagentisplunkeddowninanenvironment, itgeneratesasequenceof actionsaccordingtotheperceptsitreceives. Thissequenceofactionscausestheenvironment togothroughasequence ofstates. Ifthesequence isdesirable, thentheagenthasperformed well. This notion of desirability is captured by a performance measure that evaluates any Performance measure givensequence ofenvironment states. Humanshavedesiresandpreferences oftheirown,sothenotionofrationality asapplied to humans has to do with their success in choosing actions that produce sequences of environmentstatesthataredesirablefromtheirpointofview. Machines,ontheotherhand,donot havedesiresandpreferencesoftheirown;theperformancemeasureis,initiallyatleast,inthe mindofthedesigner ofthemachine, orinthemindoftheusers themachine isdesigned for. Wewillseethatsomeagent designs haveanexplicitrepresentation of(aversion of)theperformance measure, whileinother designs theperformance measure isentirely implicit—the agentmaydotherightthing,butitdoesn’tknowwhy. Recalling Norbert Wiener’s warning to ensure that “the purpose put into the machine is the purpose which we really desire” (page 33), notice that it can be quite hard to formulate aperformance measurecorrectly. Consider, forexample, thevacuum-cleaner agent fromthe preceding section. Wemightpropose tomeasure performance bytheamount ofdirtcleaned up in a single eight-hour shift. With a rational agent, of course, what you ask for is what you get. A rational agent can maximize this performance measure by cleaning up the dirt, then dumping it all on the floor, then cleaning it up again, and so on. A more suitable performance measure would reward the agent for having a clean floor. For example, one point could be awarded for each clean square at each time step (perhaps with a penalty for elec- ◭ tricity consumed and noise generated). As a general rule, it is better to design performance measures according to what one actually wants to be achieved in the environment, rather thanaccording tohowonethinkstheagentshouldbehave. Evenwhentheobviouspitfallsareavoided, someknottyproblemsremain. Forexample, the notion of “clean floor” in the preceding paragraph is based on average cleanliness over time. Yetthesameaveragecleanliness canbeachievedbytwodifferentagents,oneofwhich does a mediocre job all the time while the other cleans energetically but takes long breaks. Which is preferable might seem to be a fine point of janitorial science, but in fact it is a deep philosophical question with far-reaching implications. Which is better—a reckless life of highs and lows, or a safe but humdrum existence? Which is better—an economy where everyonelivesinmoderatepoverty,oroneinwhichsomeliveinplentywhileothersarevery poor? Weleavethesequestions asanexerciseforthediligent reader. For most of the book, we will assume that the performance measure can be specified correctly. Forthereasonsgivenabove,however,wemustacceptthepossibilitythatwemight put the wrong purpose into the machine—precisely the King Midas problem described on

40 Chapter 2 Intelligent Agents page 33. Moreover, when designing one piece of software, copies of which will belong to different users, wecannot anticipate the exact preferences ofeach individual user. Thus, we may need to build agents that reflect initial uncertainty about the true performance measure andlearnmoreaboutitastimegoesby;suchagentsaredescribedin Chapters16,18,and22. 2.2.2 Rationality Whatisrational atanygiventimedepends onfourthings: • Theperformance measurethatdefinesthecriterionofsuccess. • Theagent’spriorknowledge oftheenvironment. • Theactionsthattheagentcanperform. • Theagent’sperceptsequence todate. D ra e t fi io n n i a ti l o a n g o en f t a ◮ Thisleadstoadefinitionofarationalagent: Foreachpossible perceptsequence, a rationalagentshouldselect anaction thatis expectedtomaximizeitsperformancemeasure,giventheevidenceprovidedbythepercept sequenceandwhateverbuilt-inknowledgetheagenthas. Consider thesimplevacuum-cleaner agentthat cleans asquare ifitisdirty andmovestothe other square ifnot; thisistheagentfunction tabulated in Figure2.3. Isthisarational agent? That depends! First, we need to say what the performance measure is, what is known about theenvironment, andwhatsensorsandactuators theagenthas. Letusassumethefollowing: • The performance measure awards one point for each clean square at each time step, overa“lifetime” of1000timesteps. • The “geography” of the environment is known a priori (Figure 2.2) but the dirt distributionandtheinitiallocationoftheagentarenot. Cleansquaresstaycleanandsucking cleans the current square. The Right and Left actions move the agent one square except when this would take the agent outside the environment, in which case the agent remainswhereitis. • Theonlyavailable actionsare Right,Left,and Suck. • Theagentcorrectly perceives itslocation andwhetherthatlocation contains dirt. Under these circumstances the agent is indeed rational; its expected performance is at least asgoodasanyotheragent’s. Onecanseeeasilythatthesameagentwouldbeirrationalunderdifferentcircumstances. Forexample,onceallthedirtiscleanedup,theagentwilloscillateneedlesslybackandforth; iftheperformancemeasureincludesapenaltyofonepointforeachmovement,theagentwill fare poorly. A better agent for this case would do nothing once it is sure that all the squares are clean. If clean squares can become dirty again, the agent should occasionally check and re-cleanthemifneeded. Ifthegeographyoftheenvironmentisunknown,theagentwillneed toexploreit. Exercise2.VACR asksyoutodesignagentsforthesecases. 2.2.3 Omniscience, learning, and autonomy Omniscience We need to be careful to distinguish between rationality and omniscience. An omniscient agent knows the actual outcome of its actions and can act accordingly; but omniscience is impossible in reality. Consider the following example: I am walking along the Champs Elyse´es one day and I see an old friend across the street. There is no traffic nearby and I’m

Section2.2 Good Behavior:The Conceptof Rationality 41 not otherwise engaged, so, being rational, I start to cross the street. Meanwhile, at 33,000 feet, a cargo door falls off a passing airliner,3 and before I make it to the other side of the street Iamflattened. Was Iirrationaltocrossthestreet? Itisunlikelythatmyobituarywould read“Idiotattemptstocrossstreet.” Thisexampleshowsthatrationality isnotthesameasperfection. Rationalitymaximizes expected performance, while perfection maximizes actual performance. Retreating from a requirement ofperfection isnotjustaquestion ofbeingfairtoagents. Thepointisthatifwe expect an agent to dowhatturns outafter thefact tobe thebest action, it willbeimpossible todesignanagenttofulfillthisspecification—unless weimprovetheperformance ofcrystal ballsortimemachines. Our definition of rationality does not require omniscience, then, because the rational choice depends only on the percept sequence to date. We must also ensure that we haven’t inadvertently allowedtheagenttoengageindecidedly underintelligent activities. Forexample,ifanagentdoesnotlookbothwaysbeforecrossingabusyroad,thenitsperceptsequence will not tell it that there is a large truck approaching at high speed. Does our definition of rationality saythatit’snow OKtocrosstheroad? Farfromit! First,itwouldnotberationaltocrosstheroadgiventhisuninformativeperceptsequence: theriskofaccidentfromcrossingwithoutlookingistoogreat. Second,arationalagentshould choose the“looking” action before stepping into thestreet, because looking helps maximize the expected performance. Doing actions in order to modify future percepts—sometimes called information gathering—is animportant partofrationality andiscovered indepth in Information gathering Chapter 16. Asecond example ofinformation gathering is provided bythe exploration that mustbeundertaken byavacuum-cleaning agentinaninitiallyunknownenvironment. Ourdefinitionrequiresarationalagentnotonlytogatherinformationbutalsotolearnas Learning muchaspossible fromwhatitperceives. Theagent’sinitial configuration couldreflectsome prior knowledge of the environment, but asthe agent gains experience this maybe modified and augmented. There are extreme cases in which the environment is completely known a priori and completely predictable. In such cases, the agent need not perceive or learn; it simplyactscorrectly. Ofcourse,suchagentsarefragile. Considerthelowlydungbeetle. Afterdiggingitsnest and laying its eggs, it fetches a ball of dung from a nearby heap to plug the entrance. If the ballofdungisremovedfromitsgraspenroute,thebeetlecontinuesitstaskandpantomimes plugging the nest with the nonexistent dung ball, never noticing that it is missing. Evolution has built an assumption into the beetle’s behavior, and when itis violated, unsuccessful behavior results. Slightly more intelligent is the sphex wasp. The female sphex will dig a burrow, go out and sting a caterpillar and drag it to the burrow, enter the burrow again to check all is well, drag the caterpillar inside, and lay its eggs. Thecaterpillar serves as afood source when the eggs hatch. So far so good, but if an entomologist moves the caterpillar a few inches away while the sphex is doing the check, it will revert to the “drag the caterpillar” step of its plan andwillcontinuetheplanwithoutmodification,re-checkingtheburrow,evenafterdozensof caterpillar-moving interventions. The sphex is unable to learn that its innate plan is failing, andthuswillnotchange it. 3 See N.Henderson,“Newdoorlatchesurgedfor Boeing747jumbojets,”Washington Post,August24,1989.

42 Chapter 2 Intelligent Agents Totheextentthatanagentreliesonthepriorknowledgeofitsdesigner ratherthanonits Autonomy ownperceptsandlearningprocesses,wesaythattheagentlacksautonomy. Arationalagent should be autonomous—it should learn what it can to compensate for partial or incorrect prior knowledge. For example, a vacuum-cleaning agent that learns to predict where and whenadditional dirtwillappearwilldobetterthanonethatdoesnot. As a practical matter, one seldom requires complete autonomy from the start: when the agent hashadlittle ornoexperience, itwouldhave toactrandomly unless thedesigner gave some assistance. Just as evolution provides animals with enough built-in reflexes to survive longenoughtolearnforthemselves,itwouldbereasonabletoprovideanartificialintelligent agent with some initial knowledge as well as an ability to learn. After sufficient experience ofitsenvironment, thebehaviorofarationalagentcanbecomeeffectivelyindependent ofits prior knowledge. Hence, the incorporation of learning allows one todesign asingle rational agentthatwillsucceedinavastvarietyofenvironments. 2.3 The Nature of Environments Now that we have a definition of rationality, we are almost ready to think about building Taskenvironment rational agents. First, however, we must think about task environments, which are essentiallythe“problems” towhichrational agentsarethe“solutions.” Webeginbyshowinghow to specify a task environment, illustrating the process with a number of examples. We then showthattaskenvironments comeinavarietyofflavors. Thenatureofthetaskenvironment directly affectstheappropriate designfortheagentprogram. 2.3.1 Specifying the task environment In our discussion of the rationality of the simple vacuum-cleaner agent, we had to specify theperformance measure, theenvironment, andtheagent’s actuators and sensors. Wegroup allthese undertheheading ofthetaskenvironment. Fortheacronymically minded, wecall PEAS thisthe PEAS(Performance, Environment, Actuators, Sensors)description. Indesigning an agent,thefirststepmustalwaysbetospecifythetaskenvironment asfullyaspossible. The vacuum world was a simple example; let us consider a more complex problem: an automated taxi driver. Figure 2.4 summarizes the PEAS description for the taxi’s task environment. Wediscusseachelementinmoredetailinthefollowingparagraphs. First, what is the performance measure to which we would like our automated driver toaspire? Desirablequalities include gettingtothecorrectdestination; minimizingfuelconsumptionandwearandtear;minimizingthetriptimeorcost;minimizingviolationsoftraffic laws and disturbances to other drivers; maximizing safety and passenger comfort; maximizingprofits. Obviously, someofthesegoalsconflict,sotradeoffs willberequired. Next,whatisthedriving environment that thetaxiwillface? Anytaxidriver mustdeal with a variety of roads, ranging from rural lanes and urban alleys to 12-lane freeways. The roads contain other traffic, pedestrians, stray animals, road works, police cars, puddles, and potholes. The taxi must also interact with potential and actual passengers. There are also some optional choices. The taxi might need to operate in Southern California, where snow is seldom aproblem, orin Alaska, where it seldom is not. It could always be driving on the right, orwemightwantittobeflexibleenough todriveontheleftwhenin Britain or Japan. Obviously, themorerestricted theenvironment, theeasierthedesignproblem.

Section2.3 The Natureof Environments 43 Agent Type Performance Environment Actuators Sensors Measure Taxidriver Safe,fast, Roads,other Steering, Cameras,radar, legal, traffic,police, accelerator, speedometer,GPS,engine comfortable pedestrians, brake,signal, sensors,accelerometer, trip,maximize customers, horn,display, microphones,touchscreen profits, weather speech minimize impacton otherroad users Figure2.4 PEASdescriptionofthetaskenvironmentforanautomatedtaxidriver. The actuators for an automated taxi include those available to a human driver: control overtheengine through theaccelerator andcontrol oversteering andbraking. Inaddition, it will need output to a display screen or voice synthesizer to talk back to the passengers, and perhapssomewaytocommunicatewithothervehicles, politely orotherwise. Thebasicsensorsforthetaxiwillincludeoneormorevideocamerassothatitcansee,as wellas lidar and ultrasound sensors todetect distances toother cars and obstacles. Toavoid speeding tickets, the taxi should have a speedometer, and to control the vehicle properly, especially on curves, it should have an accelerometer. To determine the mechanical state of the vehicle, it will need the usual array of engine, fuel, and electrical system sensors. Like many human drivers, it might want to access GPSsignals so that it doesn’t get lost. Finally, itwillneedtouchscreen orvoiceinputforthepassenger torequestadestination. In Figure 2.5, we have sketched the basic PEAS elements for a number of additional agent types. Further examples appear in Exercise 2.PEAS. The examples include physical as well as virtual environments. Note that virtual task environments can be just as complex asthe“real” world: forexample, asoftware agent(orsoftware robot orsoftbot) that trades Softwareagent on auction and reselling Websites deals with millions of other users and billions of objects, Softbot manywithrealimages. 2.3.2 Properties of task environments The range of task environments that might arise in AI is obviously vast. We can, however, identify a fairly small number of dimensions along which task environments can be categorized. These dimensions determine, to a large extent, the appropriate agent design and the applicability of each of the principal families of techniques for agent implementation. First welistthedimensions, thenweanalyzeseveraltaskenvironments toillustrate theideas. The definitionshereareinformal;laterchaptersprovidemoreprecisestatementsandexamplesof eachkindofenvironment. Fully observable vs. partially observable: If an agent’s sensors give it access to the Fullyobservable complete state of the environment at each point in time, then we say that the task environ- Partiallyobservable ment is fully observable. A task environment is effectively fully observable if the sensors detect all aspects that are relevant tothe choice of action; relevance, inturn, depends on the

44 Chapter 2 Intelligent Agents Agent Type Performance Environment Actuators Sensors Measure Medical Healthypatient, Patient,hospital, Displayof Touchscreen/voice diagnosissystem reducedcosts staff questions,tests, entryof diagnoses, symptomsand treatments findings Satelliteimage Correct Orbitingsatellite, Displayofscene High-resolution analysissystem categorizationof downlink, categorization digitalcamera objects,terrain weather Part-picking Percentageof Conveyorbelt Jointedarmand Camera,tactile robot partsincorrect withparts;bins hand andjointangle bins sensors Refinery Purity,yield, Refinery,raw Valves,pumps, Temperature, controller safety materials, heaters,stirrers, pressure,flow, operators displays chemicalsensors Interactive Student’sscore Setofstudents, Displayof Keyboardentry, Englishtutor ontest testingagency exercises, voice feedback,speech Figure2.5 Examplesofagenttypesandtheir PEASdescriptions. performance measure. Fullyobservable environments areconvenient because theagentneed notmaintainanyinternal statetokeeptrackoftheworld. Anenvironment mightbepartially observable because of noisy and inaccurate sensors or because parts of the state are simply missing from the sensor data—for example, a vacuum agent with only a local dirt sensor cannottellwhetherthereisdirtinothersquares,andanautomatedtaxicannotseewhatother drivers are thinking. If the agent has no sensors at all then the environment is unobserv Unobservable able. One might think that in such cases the agent’s plight is hopeless, but, as wediscuss in Chapter4,theagent’s goalsmaystillbeachievable, sometimeswithcertainty. Single-agent Single-agent vs. multiagent: The distinction between single-agent and multiagent en Multiagent vironments may seem simple enough. Forexample, anagent solving acrossword puzzle by itself is clearly in a single-agent environment, whereas an agent playing chess is in a twoagent environment. However, there are some subtle issues. First, wehave described how an entity may be viewed as an agent, but we have not explained which entities must be viewed as agents. Does an agent A (the taxi driver for example) have to treat an object B (another vehicle)asanagent,orcanitbetreatedmerelyasanobjectbehavingaccordingtothelawsof physics, analogous towavesatthebeach orleaves blowing inthewind? Thekeydistinction iswhether B’sbehavior isbestdescribedasmaximizingaperformancemeasurewhosevalue depends onagent A’sbehavior.

Section2.3 The Natureof Environments 45 Forexample,inchess, theopponent entity Bistrying tomaximizeitsperformance measure, which,bytherulesofchess, minimizesagent A’sperformance measure. Thus,chessis a competitive multiagent environment. On the other hand, in the taxi-driving environment, Competitive avoiding collisions maximizes the performance measure of all agents, so it is a partially cooperativemultiagentenvironment. Itisalsopartiallycompetitivebecause,forexample,only Cooperative onecarcanoccupyaparking space. The agent-design problems in multiagent environments are often quite different from thoseinsingle-agent environments; forexample, communication oftenemerges asarational behaviorinmultiagentenvironments; insomecompetitiveenvironments, randomizedbehaviorisrational becauseitavoids thepitfallsofpredictability. Deterministic vs. nondeterministic. If the next state of the environment is completely Deterministic determined by the current state and the action executed by the agent(s), then we say the Nondeterministic environmentisdeterministic;otherwise,itisnondeterministic. Inprinciple,anagentneednot worryaboutuncertainty inafullyobservable, deterministic environment. Iftheenvironment ispartially observable, however,thenitcouldappear tobenondeterministic. Mostrealsituationsaresocomplexthatitisimpossibletokeeptrackofalltheunobserved aspects; for practical purposes, they must be treated as nondeterministic. Taxi driving is clearly nondeterministic in this sense, because one can never predict the behavior of traffic exactly; moreover, one’s tires may blow out unexpectedly and one’s engine may seize up without warning. The vacuum world as we described it is deterministic, but variations can includenondeterministic elementssuchasrandomlyappearing dirtandanunreliable suction mechanism (Exercise2.VFIN). Onefinalnote: thewordstochasticisusedbysomeasasynonymfor“nondeterministic,” Stochastic but we make a distinction between the two terms; we say that a model of the environment is stochastic if it explicitly deals with probabilities (e.g., “there’s a 25% chance of rain tomorrow”)and“nondeterministic” ifthepossibilities arelistedwithoutbeingquantified (e.g., “there’sachanceofraintomorrow”). Episodic vs. sequential: In an episodic task environment, the agent’s experience is di- Episodic vided into atomic episodes. In each episode the agent receives a percept and then performs Sequential a single action. Crucially, the next episode does not depend on the actions taken in previous episodes. Many classification tasks are episodic. For example, an agent that has to spot defective parts on an assembly line bases each decision on the current part, regardless of previous decisions; moreover, the current decision doesn’t affect whether the next part is defective. In sequential environments, on the other hand, the current decision could affect allfuture decisions.4 Chessand taxidriving aresequential: inboth cases, short-term actions can have long-term consequences. Episodic environments are much simpler than sequential environments becausetheagentdoesnotneedtothinkahead. Static vs. dynamic: If the environment can change while an agent is deliberating, then Static wesaytheenvironment isdynamic forthatagent;otherwise, itisstatic. Staticenvironments Dynamic areeasytodealwithbecausetheagentneednotkeeplookingattheworldwhileitisdeciding on an action, nor need it worry about the passage of time. Dynamic environments, on the other hand, are continuously asking the agent what it wants to do; if it hasn’t decided yet, 4 Theword“sequential”isalsousedincomputerscienceastheantonymof“parallel.” Thetwomeaningsare largelyunrelated.

46 Chapter 2 Intelligent Agents that counts as deciding to do nothing. If the environment itself does not change with the passage of time but the agent’s performance score does, then we say the environment is Semidynamic semidynamic. Taxidrivingisclearlydynamic: theothercarsandthetaxiitselfkeepmoving whilethe driving algorithm dithers about whattodonext. Chess, whenplayed withaclock, issemidynamic. Crosswordpuzzlesarestatic. Discrete Discrete vs. continuous: The discrete/continuous distinction applies to the state of the Continuous environment, to the way time is handled, and to the percepts and actions of the agent. For example, the chess environment has a finite number of distinct states (excluding the clock). Chess also has a discrete set of percepts and actions. Taxi driving is a continuous-state and continuous-time problem: the speed and location ofthe taxi and ofthe other vehicles sweep through arangeofcontinuous values anddososmoothlyovertime. Taxi-driving actions are also continuous (steering angles, etc.). Input from digital cameras isdiscrete, strictly speaking,butistypically treatedasrepresenting continuously varyingintensities andlocations. Known Known vs. unknown: Strictly speaking, this distinction refers not to the environment Unknown itself but to the agent’s (or designer’s) state of knowledge about the “laws of physics” of the environment. In a known environment, the outcomes (or outcome probabilities if the environment is nondeterministic) for all actions are given. Obviously, if the environment is unknown, theagentwillhavetolearnhowitworksinordertomakegooddecisions. The distinction between known and unknown environments is not the same as the one betweenfullyandpartiallyobservableenvironments. Itisquitepossibleforaknownenvironment to be partially observable—for example, in solitaire card games, I know the rules but am still unable to see the cards that have not yet been turned over. Conversely, an unknown environment can be fully observable—in a new video game, the screen may show the entire gamestatebut Istilldon’tknowwhatthebuttonsdountil Itrythem. As noted on page 39, the performance measure itself may be unknown, either because the designer is not sure how to write it down correctly or because the ultimate user—whose preferences matter—is not known. Forexample, ataxi driver usually won’t know whether a new passenger prefers a leisurely or speedy journey, a cautious or aggressive driving style. A virtual personal assistant starts out knowing nothing about the personal preferences of its newowner. Insuchcases,theagentmaylearnmoreabouttheperformancemeasurebasedon furtherinteractionswiththedesigneroruser. This,inturn,suggeststhatthetaskenvironment isnecessarily viewedasamultiagentenvironment. The hardest case is partially observable, multiagent, nondeterministic, sequential, dynamic, continuous, and unknown. Taxi driving is hard in all these senses, except that the driver’senvironment ismostlyknown. Drivingarentedcarinanewcountry withunfamiliar geography, differenttrafficlaws,andnervouspassengers isalotmoreexciting. Figure2.6 lists theproperties of anumber offamiliar environments. Note thatthe properties are not always cut and dried. For example, we have listed the medical-diagnosis task assingle-agent becausethediseaseprocessinapatientisnotprofitablymodeledasanagent; but amedical-diagnosis system might also haveto dealwithrecalcitrant patients andskepticalstaff,sotheenvironment couldhaveamultiagentaspect. Furthermore, medicaldiagnosis isepisodic ifone conceives ofthetask asselecting adiagnosis given alistofsymptoms; the problem is sequential if the task can include proposing a series of tests, evaluating progress overthecourseoftreatment, handling multiplepatients, andsoon.

Section2.1 Agentsand Environments 37 Agent Sensors Actuators Environment Percepts ? Actions Figure2.1 Agentsinteractwithenvironmentsthroughsensorsandactuators. We can imagine tabulating the agent function that describes any given agent; for most agents, this would be a very large table—infinite, in fact, unless we place a bound on the lengthofperceptsequenceswewanttoconsider. Givenanagenttoexperimentwith,wecan, in principle, construct this table by trying out all possible percept sequences and recording whichactionstheagentdoesinresponse.1 Thetableis,ofcourse,anexternalcharacterization of the agent. Internally, the agent function for an artificial agent will be implemented by an agent program. It is important to keep these two ideas distinct. The agent function is an Agentprogram abstract mathematical description; the agent program is a concrete implementation, running withinsomephysicalsystem. To illustrate these ideas, we use a simple example—the vacuum-cleaner world, which consists of a robotic vacuum-cleaning agent in a world consisting of squares that can be either dirty or clean. Figure 2.2 shows a configuration with just two squares, A and B. The vacuum agent perceives which square it is in and whether there is dirt in the square. The agentstartsinsquare A. Theavailable actionsaretomovetotheright,movetotheleft,suck up the dirt, or do nothing.2 One very simple agent function is the following: if the current square is dirty, then suck; otherwise, move to the other square. A partial tabulation of this agent function is shown in Figure 2.3 and an agent program that implements it appears in Figure2.8onpage49. Looking at Figure 2.3, we see that various vacuum-world agents can be defined simply ◭ byfillingintheright-handcolumninvariousways. Theobviousquestion,then,isthis: What is the right way to fill out the table? In other words, what makes an agent good or bad, intelligent orstupid? Weanswerthesequestions inthenextsection. 1 Iftheagentusessomerandomizationtochoose itsactions, thenwewouldhavetotryeachsequence many timestoidentifytheprobabilityofeachaction. Onemightimaginethatactingrandomlyisrathersilly,butwe showlaterinthischapterthatitcanbeveryintelligent. 2 Inarealrobot,itwouldbeunlikelytohaveanactionslike“moveright”and“moveleft.” Insteadtheactions wouldbe“spinwheelsforward”and“spinwheelsbackward.” Wehavechosentheactionstobeeasiertofollow onthepage,notforeaseofimplementationinanactualrobot.

50 Chapter 2 Intelligent Agents Agent Environment Sensors What the world is like now What action I Condition-action rules should do now Actuators Figure 2.9 Schematic diagram of a simple reflex agent. We use rectangles to denote the currentinternalstateoftheagent’sdecisionprocess,andovalstorepresentthebackground informationusedintheprocess. visualinputtoestablishtheconditionwecall“Thecarinfrontisbraking.” Then,thistriggers some established connection in the agent program to the action “initiate braking.” We call Condition–action suchaconnection acondition–action rule,6 writtenas rule ifcar-in-front-is-braking theninitiate-braking. Humansalsohavemanysuchconnections, someofwhicharelearned responses(asfordriving)andsomeofwhichareinnatereflexes(suchasblinkingwhensomething approaches the eye). In the course of the book, we show several different ways in which such connections canbelearnedandimplemented. The program in Figure 2.8 is specific to one particular vacuum environment. A more general and flexible approach is first to build a general-purpose interpreter for condition– action rules and then to create rule sets for specific task environments. Figure 2.9 gives the structure ofthisgeneral programinschematicform,showinghowthecondition–action rules allow the agent to make the connection from percept to action. Do not worry if this seems trivial;itgetsmoreinteresting shortly. An agent program for Figure 2.9 is shown in Figure 2.10. The INTERPRET-INPUT function generates an abstracted description of the current state from the percept, and the RULE-MATCH function returns the first rule in the set of rules that matches the given state description. Note that the description in terms of “rules” and “matching” is purely conceptual; as noted above, actual implementations can be as simple as a collection of logic gates implementinga Booleancircuit. Alternatively,a“neural”circuitcanbeused,wherethelogic gatesarereplaced bythenonlinear unitsofartificialneuralnetworks(see Chapter21). ◮ Simplereflexagentshavetheadmirableproperty ofbeingsimple,buttheyareoflimited intelligence. Theagent in Figure 2.10willwork only ifthecorrect decision canbe madeon thebasisofjustthecurrentpercept—that is,onlyiftheenvironment isfullyobservable. 6 Alsocalledsituation–actionrules,productions,orif–thenrules.

Section2.4 The Structureof Agents 49 function REFLEX-VACUUM-AGENT([location,status])returnsanaction ifstatus=Dirtythenreturn Suck elseiflocation=Athenreturn Right elseiflocation=Bthenreturn Left Figure2.8 Theagentprogramforasimplereflexagentinthetwo-locationvacuumenvironment. Thisprogramimplementstheagentfunctiontabulatedin Figure2.3. Thekeychallenge for AIistofindouthowtowriteprograms that, totheextent possible, produce rationalbehavior fromasmallishprogram ratherthanfromavasttable. We have many examples showing that this can be done successfully in other areas: for example, the huge tables of square roots used by engineers and schoolchildren prior to the 1970s havenowbeen replaced byafive-lineprogram for Newton’smethodrunning onelectronic calculators. The question is, can AI do for general intelligent behavior what Newton didforsquareroots? Webelievetheanswerisyes. In the remainder of this section, we outline four basic kinds of agent programs that embodytheprinciples underlying almostallintelligent systems: • Simplereflexagents; • Model-based reflexagents; • Goal-basedagents; and • Utility-based agents. Each kind of agent program combines particular components in particular ways to generate actions. Section2.4.6explains ingeneral termshowtoconvert alltheseagents intolearning agentsthatcanimprovetheperformanceoftheircomponentssoastogeneratebetteractions. Finally, Section2.4.7describes thevariety ofwaysinwhichthecomponents themselves can be represented within the agent. This variety provides a major organizing principle for the fieldandforthebookitself. 2.4.2 Simple reflex agents Thesimplestkindofagentisthesimplereflexagent. Theseagentsselectactionsonthebasis Simplereflexagent ofthecurrentpercept,ignoringtherestofthepercepthistory. Forexample,thevacuumagent whose agent function is tabulated in Figure 2.3is asimple reflexagent, because its decision is based only on the current location and on whether that location contains dirt. An agent program forthisagentisshownin Figure2.8. Noticethatthevacuum agentprogram isverysmallindeed compared tothecorresponding table. The most obvious reduction comes from ignoring the percept history, which cuts down the number of relevant percept sequences from 4T to just 4. A further, small reduction comesfrom thefact thatwhenthe current square isdirty, theaction does notdepend on thelocation. Although wehavewritten theagentprogram using if-then-else statements, itis simpleenough thatitcanalsobeimplementedasa Booleancircuit. Simplereflexbehaviors occur eveninmorecomplex environments. Imagine yourself as the driver of the automated taxi. If the car in front brakes and its brake lights come on, then you should notice this and initiate braking. In other words, some processing is done on the

Section2.4 The Structureof Agents 47 Task Environment Observable Agents Deterministic Episodic Static Discrete Crosswordpuzzle Fully Single Deterministic Sequential Static Discrete Chesswithaclock Fully Multi Deterministic Sequential Semi Discrete Poker Partially Multi Stochastic Sequential Static Discrete Backgammon Fully Multi Stochastic Sequential Static Discrete Taxidriving Partially Multi Stochastic Sequential Dynamic Continuous Medicaldiagnosis Partially Single Stochastic Sequential Dynamic Continuous Imageanalysis Fully Single Deterministic Episodic Semi Continuous Part-pickingrobot Partially Single Stochastic Episodic Dynamic Continuous Refinerycontroller Partially Single Stochastic Sequential Dynamic Continuous Englishtutor Partially Multi Stochastic Sequential Dynamic Discrete Figure2.6 Examplesoftaskenvironmentsandtheircharacteristics. We have not included a “known/unknown” column because, as explained earlier, this is not strictly aproperty of the environment. Forsome environments, such aschess and poker, it is quite easy to supply the agent with full knowledge of the rules, but it is nonetheless interestingtoconsiderhowanagentmightlearntoplaythesegameswithoutsuchknowledge. Thecoderepository associatedwiththisbook(aima.cs.berkeley.edu)includesmultiple environment implementations, together with a general-purpose environment simulator for evaluating an agent’s performance. Experiments are often carried out not for a single environment butformanyenvironments drawnfrom anenvironmentclass. Forexample, to Environment class evaluate a taxi driver in simulated traffic, we would want to run many simulations with different traffic, lighting, and weather conditions. Weare then interested inthe agent’s average performance overtheenvironment class. 2.4 The Structure of Agents Sofarwehavetalkedaboutagentsbydescribingbehavior—theactionthatisperformedafter any given sequence ofpercepts. Nowwemust bite the bullet and talk about how theinsides work. The job of AI is to design an agent program that implements the agent function— Agentprogram the mapping from percepts to actions. We assume this program will run on some sort of computing devicewithphysicalsensorsandactuators—we callthistheagentarchitecture: Agentarchitecture agent=architecture+program. Obviously,theprogramwechoosehastobeonethatisappropriateforthearchitecture. Ifthe program isgoing torecommend actions like Walk,thearchitecture hadbetterhave legs. The architecture might be just an ordinary PC, or it might be a robotic car with several onboard computers, cameras, and other sensors. In general, the architecture makes the percepts from thesensorsavailabletotheprogram,runstheprogram,andfeedstheprogram’sactionchoices to the actuators asthey are generated. Most ofthis book is about designing agent programs, although Chapters25and26dealdirectly withthesensorsandactuators.

AGENT, Black's Law Dictionary (11th ed. 2019) - stock-transfer agent. (1873) See transfer agent. - subagent. (18c) 1. A person to whom an agent has delegated the performance of an act for the principal; a person designated by an agent to perform some duty relating to the agency. • If the principal consents to a primary agent's employment of a subagent, the subagent owes fiduciary duties to the principal, and the principal is liable for the subagent's acts. — Also termed subservant. Cf. primary agent; subordinate agent. “By delegation … the agent is permitted to use agents of his own in performing the function he is employed to perform for his principal, delegating to them the discretion which normally he would be expected to exercise personally. These agents are known as subagents to indicate that they are the agent's agents and not the agents of the principal. Normally (though of course not necessarily) they are paid by the agent. The agent is liable to the principal for any injury done him by the misbehavior of the agent's subagents.” Floyd R. Mechem, Outlines of the Law of Agency § 79, at 51 (Philip Mechem ed., 4th ed. 1952). 2. See buyer's broker under broker. - subordinate agent. (17c) An agent who acts subject to the direction of a superior agent. • Subordinate and superior agents are co-agents of a common principal. See superior agent. Cf. subagent (1). - successor agent. (1934) An agent who is appointed by a principal to act in a primary agent's stead if the primary agent is unable or unwilling to perform. - superior agent. (17c) 1. An agent on whom a principal confers the right to direct a subordinate agent. See subordinate agent. 2. See high-managerial agent (1). - transfer agent. (1850) An organization (such as a bank or trust company) that handles transfers of shares for a publicly held corporation by issuing new certificates and overseeing the cancellation of old ones and that usu. also maintains the record of shareholders for the corporation and mails dividend checks. • Generally, a transfer agent ensures that certificates submitted for transfer are properly indorsed and that the transfer right is appropriately documented. — Also termed stock-transfer agent. - trustee-agent. A trustee who is subject to the control of the settlor or one or more beneficiaries of a trust. See trustee (1). - undercover agent. (1930) 1. An agent who does not disclose his or her role as an agent. 2. A police officer who gathers evidence of criminal activity without disclosing his or her identity to the suspect. - undisclosed agent. (1863) An agent who deals with a third party who has no knowledge that the agent is acting on a principal's behalf. Cf. undisclosed principal under principal (1). - universal agent. (18c) An agent authorized to perform all acts that the principal could personally perform. - vice-commercial agent. (1800) Hist. In the consular service of the United States, a consular officer who was substituted temporarily to fill the place of a commercial agent who was absent or had been relieved from duty. Westlaw. © 2019 Thomson Reuters. No Claim to Orig. U.S. Govt. Works. End of Document © 2024 Thomson Reuters. No claim to original U.S. Government Works. © 2024 Thomson Reuters. No claim to original U.S. Government Works. 4

AGENT, Black's Law Dictionary (11th ed. 2019) Black's Law Dictionary (11th ed. 2019), agent AGENT Bryan A. Garner, Editor in Chief Preface | Guide | Legal Maxims | Bibliography agent (15c) 1. Something that produces an effect <an intervening agent>. See cause (1); electronic agent. 2. Someone who is authorized to act for or in place of another; a representative <a professional athlete's agent>. — Also termed commissionaire. See agency. Cf. principal, n. (1); employee. “Generally speaking, anyone can be an agent who is in fact capable of performing the functions involved. The agent normally binds not himself but his principal by the contracts he makes; it is therefore not essential that he be legally capable to contract (although his duties and liabilities to his principal might be affected by his status). Thus an infant or a lunatic may be an agent, though doubtless the court would disregard either's attempt to act if he were so young or so hopelessly devoid of reason as to be completely incapable of grasping the function he was attempting to perform.” Floyd R. Mechem, Outlines of the Law of Agency 8–9 (Philip Mechem ed., 4th ed. 1952). “The etymology of the word agent or agency tells us much. The words are derived from the Latin verb, ago, agere; the noun agens, agentis. The word agent denotes one who acts, a doer, force or power that accomplishes things.” Harold Gill Reuschlein & William A. Gregory, The Law of Agency and Partnership § 1, at 2–3 (2d ed. 1990). - agent not recognized. Patents. A patent applicant's appointed agent who is not registered to practice before the U.S. Patent and Trademark Office. • A power of attorney appointing an unregistered agent is void. See patent agent. - agent of necessity. (1857) An agent that the law empowers to act for the benefit of another in an emergency. — Also termed agent by necessity. - apparent agent. (1823) Someone who reasonably appears to have authority to act for another, regardless of whether actual authority has been conferred. — Also termed ostensible agent; implied agent. - associate agent. Patents. An agent who is registered to practice before the U.S. Patent and Trademark Office, has been appointed by a primary agent, and is authorized to prosecute a patent application through the filing of a power of attorney. • An associate agent is often used by outside counsel to assist in-house counsel. See patent agent. - bail-enforcement agent. See bounty hunter. - bargaining agent. (1935) A labor union in its capacity of representing employees in collective bargaining. - broker-agent. See broker. - business agent. See business agent. - case agent. See case agent. - clearing agent. (1937) Securities. A person or company acting as an intermediary in a securities transaction or providing facilities for comparing data regarding securities transactions. • The term includes a custodian of securities in connection with the central handling of securities. Securities Exchange Act § 3(a)(23)(A) (15 USCA § 78c(a)(23)(A)). — Also termed clearing agency. - closing agent. (1922) An agent who represents the purchaser or buyer in the negotiation and closing of a real-property transaction by handling financial calculations and transfers of documents. — Also termed settlement agent. See also settlement attorney under attorney. - co-agent. (16c) Someone who shares with another agent the authority to act for the principal. — Also termed dual agent. Cf. common agent. - commercial agent. (18c) 1. broker. 2. A consular officer responsible for the commercial interests of his or her country at a foreign port. 3. See mercantile agent. 4. See commission agent. - commission agent. (1812) An agent whose remuneration is based at least in part on commissions, or percentages of actual sales. • Commission agents typically work as middlemen between sellers and buyers. — Also termed commercial agent. © 2024 Thomson Reuters. No claim to original U.S. Government Works. 1

AGENT, Black's Law Dictionary (11th ed. 2019) - common agent. (17c) An agent who acts on behalf of more than one principal in a transaction. Cf. co-agent. - corporate agent. (1819) An agent authorized to act on behalf of a corporation; broadly, all employees and officers who have the power to bind the corporation. - county agent. See juvenile officer under officer (1). - del credere agent (del kred-ə-ray or kray-də-ray) (1822) An agent who guarantees the solvency of the third party with whom the agent makes a contract for the principal. • A del credere agent receives possession of the principal's goods for purposes of sale and guarantees that anyone to whom the agent sells the goods on credit will pay promptly for them. For this guaranty, the agent receives a higher commission for sales. The promise of such an agent is almost universally held not to be within the statute of frauds. — Also termed del credere factor. - diplomatic agent. (18c) A national representative in one of four categories: (1) ambassadors, (2) envoys and ministers plenipotentiary, (3) ministers resident accredited to the sovereign, or (4) chargés d'affaires accredited to the minister of foreign affairs. - double agent. (1935) 1. A spy who finds out an enemy's secrets for his or her principal but who also gives secrets to the enemy. 2. See dual agent (2). - dual agent. (1881) 1. See co-agent. 2. An agent who represents both parties in a single transaction, esp. a buyer and a seller. — Also termed (in sense 2) double agent. - emigrant agent. (1874) One engaged in the business of hiring laborers for work outside the country or state. - enrolled agent. See enrolled agent. - escrow agent. See escrow agent. - estate agent. See real-estate agent. - fiscal agent. (18c) A bank or other financial institution that collects and disburses money and services as a depository of private and public funds on another's behalf. - foreign agent. (1938) Someone who registers with the federal government as a lobbyist representing the interests of a foreign country or corporation. - forwarding agent. (1837) 1. freight forwarder. 2. A freight-forwarder who assembles less-than-carload shipments (small shipments) into carload shipments, thus taking advantage of lower freight rates. - general agent. (17c) An agent authorized to transact all the principal's business of a particular kind or in a particular place. • Among the common types of general agents are factors, brokers, and partners. Cf. special agent. - government agent. (1805) 1. An employee or representative of a governmental body. 2. A law-enforcement official, such as a police officer or an FBI agent. 3. An informant, esp. an inmate, used by law enforcement to obtain incriminating statements from another inmate. - gratuitous agent. (1822) An agent who acts without a right to compensation. - high-managerial agent. (1957) 1. An agent of a corporation or other business who has authority to formulate corporate policy or supervise employees. — Also termed superior agent. 2. See superior agent (1). - implied agent. See apparent agent. - independent agent. (17c) An agent who exercises personal judgment and is subject to the principal only for the results of the work performed. Cf. nonservant agent. - innocent agent. (1805) Criminal law. A person whose action on behalf of a principal is unlawful but does not merit prosecution because the agent had no knowledge of the principal's illegal purpose; a person who lacks the mens rea for an offense but who is tricked or coerced by the principal into committing a crime. • Although the agent's conduct was unlawful, the agent might not be prosecuted if the agent had no knowledge of the principal's illegal purpose. The principal is legally accountable for the innocent agent's actions. See Model Penal Code § 2.06(2)(a). - insurance agent. See insurance agent. - jural agent. See jural agent. - land agent. See land agent. - listing agent. (1927) The real-estate broker's representative who obtains a listing agreement with the owner. Cf. selling agent; showing agent. - local agent. (1804) 1. An agent appointed to act as another's (esp. a company's) representative and to transact business within a specified district. 2. See special agent. © 2024 Thomson Reuters. No claim to original U.S. Government Works. 2

AGENT, Black's Law Dictionary (11th ed. 2019) - managing agent. (1812) A person with general power involving the exercise of judgment and discretion, as opposed to an ordinary agent who acts under the direction and control of the principal. — Also termed business agent. - mercantile agent. (18c) An agent employed to sell goods or merchandise on behalf of the principal. — Also termed commercial agent. - nonservant agent. (1920) An agent who agrees to act on the principal's behalf but is not subject to the principal's control over how the task is performed. • A principal is not liable for the physical torts of a nonservant agent. See independent contractor. Cf. independent agent; servant. - ostensible agent. See apparent agent. - patent agent. (1859) A specialized legal professional — not necessarily a lawyer — who has fulfilled the U.S. Patent and Trademark Office requirements as a representative and is registered to prepare and prosecute patent applications before the PTO. • To be registered to practice before the PTO, a candidate must establish mastery of the relevant technology (by holding a specified technical degree or equivalent training) in order to advise and assist patent applicants. The candidate must also pass a written examination (the “Patent Bar”) that tests knowledge of patent law and PTO procedure. — Often shortened to agent. — Also termed registered patent agent; patent solicitor. Cf. patent attorney. - primary agent. (18c) An agent who is directly authorized by a principal. • A primary agent generally may hire a subagent to perform all or part of the agency. Cf. subagent (1). - private agent. (17c) An agent acting for an individual in that person's private affairs. - process agent. (1886) A person authorized to accept service of process on behalf of another. See registered agent. - procuring agent. (1954) Someone who obtains drugs on behalf of another person and delivers the drugs to that person. • In criminal-defense theory, the procuring agent does not sell, barter, exchange, or make a gift of the drugs to the other person because the drugs already belong to that person, who merely employs the agent to pick up and deliver them. - public agent. (17c) A person appointed to act for the public in matters relating to governmental administration or public business. - real-estate agent. (1844) An agent who represents a buyer or seller (or both, with proper disclosures) in the sale or lease of real property. • A real-estate agent can be either a broker (whose principal is a buyer or seller) or a salesperson (whose principal is a broker). — Also termed estate agent. Cf. realtor. - record agent. See insurance agent. - registered agent. (1809) A person authorized to accept service of process for another person, esp. a foreign corporation, in a particular jurisdiction. — Also termed resident agent. See process agent. - registered patent agent. See patent agent. - resident agent. See registered agent. - secret agent. See secret agent. - selling agent. (1839) 1. The real-estate broker's representative who sells the property, as opposed to the agent who lists the property for sale. 2. See showing agent. Cf. listing agent. - settlement agent. (1952) See closing agent. - showing agent. (1901) A real-estate broker's representative who markets property to a prospective purchaser. • A showing agent may be characterized as a subagent of the listing broker, as an agent who represents the purchaser, or as an intermediary who owes an agent's duties to neither seller nor buyer. — Also termed selling agent. Cf. listing agent. - soliciting agent. (1855) 1. Insurance. An agent with authority relating to the solicitation or submission of applications to an insurance company but usu. without authority to bind the insurer, as by accepting the applications on behalf of the company. 2. An agent who solicits orders for goods or services for a principal. 3. A managing agent of a corporation for purposes of service of process. - special agent. (17c) 1. An agent employed to conduct a particular transaction or to perform a specified act. Cf. general agent. 2. See insurance agent. - specially accredited agent. (1888) An agent that the principal has specially invited a third party to deal with, in an implication that the third party will be notified if the agent's authority is altered or revoked. - statutory agent. (1844) An agent designated by law to receive litigation documents and other legal notices for a nonresident corporation. • In most states, the secretary of state is the statutory agent for such corporations. Cf. agency by operation of law (1) under agency (1). © 2024 Thomson Reuters. No claim to original U.S. Government Works. 3

Intelligent agents: theory and practice 125 would apply its inference rules wherever possible, in order to generate the deductive closure of its base beliefs under its deduction rules. We model deductive closure in a function close: where 6.1--,, rp means that rpcan be proved from 6. using only the rules in p. A belief logic can then be defined, with the semantics to a modal belief connective [i], where i is an agent, given in terms of the deduction structured; modelling i's belief system: [i]qi iff rp e c!ose(d;). Konolige went on to examine the properties of the deduction model at some length, and developed a variety of proof methods for his logics, including resolution and tableau systems (Geissler & Konolige, 1986). The deduction model is undoubtedly simple; however, as a direct model of the belief systems of AI agents, it has much to commend it. 2.4.3 Meta-languages and syntactic modalities A meta-language is one in which it is possible to represent the properties of another language. A first-order meta-language is a first-order logic, with the standard predicates, quantifier, terms, and so on, whose domain contains formulae of some other language, called the object language. Using a meta-language, it is possible to represent a relationship between a meta-language term denoting an agent, and an object language term denoting some formula. For example, the meta-language formula Bel(Janine,[Father(Zeus, Cronos)]) might be used to represent the example (1) that we saw earlier. The quote marks, [ ... ], are used to indicate that their contents are a meta-language term denoting the corresponding object-language formula. Unfortunately, meta-language formalisms have their own package of problems, not the least of which is that they tend to fall prey to inconsistency (Montague, 1963; Thomason, 1980). However, there have been some fairly successful meta-language formalisms, including those by Konolige (1982), Haas (1986), Morgenstern (1987), and Davies (1993). Some results on retrieving consist ency appeared in the late 1980s (Pcrlis, 1985, 1988; des Rivieres & Levesque, 1986; Turner, 1990). 2.5 Pro-attitudes: goals and desires An obvious approach to developing a logic of goals or desires is to adapt possible worlds semantics-sec, e.g .. Cohen and Levesque (1990a), Wooldridge (1994). In this view, each goal accessible world represents one way the world might be if the agent's goals were realised. However, this approach falls prey to the side effect problem, in that it predicts that agents have a goal of the logical consequences of their goals (cf. the logical omniscience problem, discussed above). This is not a desirable property: one might have a goal of going to the dentist, with the necessary consequence of suffering pain, without having a goal of suffering pain. The problem is discussed (in the context of intentions), in Bratman ( 1990). The basic possible worlds model has been adapted by some researchers in an attempt to overcome this problem (Wainer, 1994). Other, related semantics for goals have been proposed (Doyle et al., 1991; Kiss & Reichgelt, 1992; Rao& Georgeff, 1991b). 2.6 Theories of agency All of the formalisms considered so far have focused on just one aspect of agency. However, it is to be expected that a realistic agent theory will be represented in a logical framework that combines these various components. Additionally, we expect an agent logic to be capable of representing the dynamic aspects of agency. A complete agent theory, expressed in a logic with these properties, must define how the attributes of agency are related. For example, it will need to show how an agent's information and pro-attitudes are related; how an agent's cognitive state changes over time; how the environment affects an agent's cognitive state; and how an agent's information and pro attitudes lead it to perform actions. Giving a good account of these relationships is the most significant problem faced by agent theorists.

Intelligent agents: theory and practice 129 theories as specifications, and agent logics as specification languages, is that the problems and issues we then face are familiar from the discipline of software engineering: How useful or expressive is the specification language? How concise are agent specifications? How does one refine or otherwise transform a specification into an implementation? However, the view of agent theories as specifications is not shared by all researchers. Some intend their agent theories to be used as knowledge representation formalisms, which raises the difficult problem of algorithms to reason with such theories. Still others intend their work to formalise a concept of interest in cognitive science or philosophy (this is, of course, what Hintikka intended in his early work on logics of knowledge of belief). What is clear is that it is important to be precise about the role one expects an agent theory to play. 2. 9 Further reading For a recent discussion on the role of logic and agency, which lays out in more detail some contrasting views on the subject, see Israel (1993, pp. 17-24). For a detailed discussion of intentionality and the intentional stance, see Dennett (1978, 1987). A number of papers on AI treatments of agency may be found in Allen et al. ( 1990). For an introduction to modal logic, sec Chellas (1980); a slightly older, though more wide ranging introduction, may be found in Hughes and Cresswell (1968). As for the use of modal logics to model knowledge and belief, see Halpern and Moses (1992), which includes complexity results and proof procedures. Related work on modelling knowledge has been done by the distributed systems community, who give the worlds in possible worlds semantics a precise interpretation; for an introduction and further references, see Halpern (1987) and Fagin et al. (1992). Overviews of formalisms for modelling belief and knowledge may be found in Halpern (1986), Konolige (1986a), Reichgelt (1989a) and Wooldridge (1992). A variant of the possible worlds framework, called the recursive modelling method, is described in Gmytrasiewicz and Durfee (1993); a deep theory of belief may be found in Mack (1994). Situation semantics, developed in the early 1980s and recently the subject of renewed interest, represent a fundamentally new approach to modelling the world and cognitive systems (Barwise & Perry, 1983; Devlin, 1991). However, situation semantics are not (yet) in the mainstream of (D)AJ, and it is not obvious what impact the paradigm will ultimately have. Logics which integrate time with mental states are discussed in Kraus and Lehmann (1988), Halpern and Vardi (1989) and Wooldridge and Fisher (1994); the last of these presents a tableau based proof method for a temporal belief logic. Two other important references for temporal aspects are Shoham (1988. 1989). Thomas has developed some logics. for representing agent theories as part of her framework for agent programming languages; see Thomas et al. (1991) and Thomas (1993) and section 4. For an introduction to temporal logics and related topics, see Goldblatt (1987) and Emerson (1990). A non-formal discussion of intention may be found in Bratman (1987), or more briefly (Bratman, 1990). Further work on modelling intention may be found in Grosz and Sidner (1990), Sadek (1992), Goldman and Lang (1991), Konolige and Pollack (1993), Bell (1995) and Dongha (1995). Related work, focusing less on single-agent attitudes, and more on social aspects, is Levesque et al. (1990), Jennings (1993a), Wooldridge (1994) and Wooldridge and Jennings (1994). Finally, although we have not discussed formalisms for reasoning about action here, we suggested above that an agent logic would need to incorporate some mechanism for representing agent's actions. Our reason for avoiding the topic is simply that the field is so big, it deserves a whole review in its own right. Good starting points for AI treatments of action arc Allen (1984), and Allen et al. (1990, 1991). Other treatments of action in agent logics arc based on formalisms borrowed from mainstream computer science, notably dynamic logic (originally developed to reason about computer programs) (Hare!, 1984). The logic of seeing to it that has been discussed in the formal philosophy literature, but has yet to impact on (D)AI (Belnap & Perloff, 1988; Perloff, 1991; Belnap, 1991; Segerberg, 1989).

M. WOOLDRIDGE AND NICHOLAS JENNINGS 128 interchange format (KIF). KOML provides the agent designer with a standard syntax for messages, and a number of performatives that define the force of a message. Example performatives include tell, perform, and reply; the inspiration for these message types comes largely from speech act theory. KIF provides a syntax for message content-KIF is essentially the first-order predicate calculus, recast in a LISP-like syntax. 2.8 Discussion Formalisms for reasoning about agents have come a long way since Hintikka's pioneering work on logics of knowledge and belief (Hintikka, 1962). Within AI, perhaps the main emphasis of subsequent work has been on attempting to develop formalisms that capture the relationship between the various elements that comprise an agent's cognitive state; the paradigm example of this work is the well-known theory of intention developed by Cohen and Levesque (1990a). Despite the very real progress that has been made, there still remain many fairly fundamental problems and issues still outstanding. On a technical level, we can identify a number of issues that remain open. First, the problems associated with possible worlds semantics (notably, logical omniscience) cannot be regarded as solved. As we observed above, possible worlds remain the semantics of choice for many researchers, and yet they do not in general represent a realistic model of agents with limited resources-and of course all real agents are resource-bounded. One solution is to ground possible worlds semantics, giving them a precise interpretation in terms of the world. This was the approach taken in Rosenschein and Kaelbling's situated automata paradigm, and can be very successful. However, it is not clear how such a grounding could be given to proattitudes such as desires or intentions (although some attempts have been made (Singh, 1990a; Wooldridge, 1992; Werner, 1990)). There is obviously much work remaining to be done on formalisms for knowledge and belief, in particular in the area of modelling resource bounded reasoners. With respect to logics that combine different attitudes, perhaps the most important problems still outstanding relate to intention. In particular, the relationship between intention and action has not been formally represented in a satisfactory way The problem seems to be that having an intention to act makes it more likely that an agent will act, but does not generally guarantee it. While it seems straightforward to build systems that appear to have intentions (Wooldridge, 1995), it seems much harder to capture this relationship formally. Other problems that have not yet really been addressed in the literature include the management of multiple, possibly conflicting intentions, and the formation, scheduling, and reconsideration of intentions. The question of exactly which combination of attitudes is required to characterise an agent is also the subject of some debate. As we observed above, a currently popular approach is to use a combination of beliefs, desires, and intentions (hence BDI architectures (Rao and Georgeff, 199lb)). However, there are alternatives: Shoham, for example, suggests that the notion of choice is more fundamental (Shoham, 1990). Comparatively little work has yet been done on formally comparing the suitability of these various combinations. One might draw a parallel with the use of temporal logics in mainstream computer science, where the expressiveness of specification languages is by now a well-understood research area (Emerson & Halpern, 1986). Perhaps the obvious requirement for the short term is experimentation with real agent specifications, in order to gain a better understanding of the relative merits of different formalisms. More general!y, the kinds of logics used in agent theory tend to be rather elaborate, typically containing many modalities which interact with each other in subtle ways. Very little work has yet been carried out on the theory underlying such logics (perhaps the only notable exception is Catach, 1988). Until the general principles and limitations of such multi-modal logics become understood, we might expect that progress with using such logics will be slow. One area in which work is likely to be done in the near future is theorem proving techniques for multi-modal logics. Finally, there is often some confusion about the role played by a theory of agency. The view we take is that such theories represent specifications for agents. The advantage of treating agent

Intelligent agents: theory and practice 127 were used: beliefs and goals. Further attitudes, such as intention, were defined in terms of these. In related work, Rao and Georgeff have developed a logical framework for agent theory based on three primitive modalities: beliefs, desires and intentions (Rao & Georgeff, 1991a,b, 1993). Their formalism is based on a branching model of time (cf. Emerson & Halpern, 1986), in which belief-, desire- and intention-accessible worlds are themselves branching time structures. They are particularly concerned with the notion of realism-the question of how an agent's beliefs about the future affect its desires and intentions. In other work, they also consider the potential for adding (social) plans to their formalism (Rao & Georgcff, 1992b; Kinny et al., 1992). 2.6.4 Singh A quite different approach to modelling agents was taken by Singh, who has developed an interesting family of logics for representing intentions, beliefs, knowledge, know-how, and communication in a branching-time framework (Singh, 1990, 199la,b; Singh & Asher, 1991); these articles are collected and expanded in Singh (1994). Singh's formalism is extremely rich, and considerable effort has been devoted to establishing its properties. However, its complexity prevents a detailed discussion here. 2.6.5 Werner In an extensive sequence of papers, Werner has laid the foundations of a general model of agency, which draws upon work in economics, game theory, situated automata theory, situation semantics, and philosophy (Werner, 1988, 1989, 1990, 1991). At the time of writing, however, the properties of this model have not been investigated in depth. 2.6.6 Wooldridge-modelling multi-agent systems For his 1992 doctoral thesis, Wooldridge developed a family of logics for representing the properties of multi-agent systems (Wooldridge, 1992; Wooldridge & Fisher, 1992). Unlike the approaches cited above, Wooldridge's aim was not to develop a general framework for agent theory. Rather, he hoped to construct formalisms that might be used in the specification and verification of realistic multi-agent systems. To this end, he developed a simple, and in some sense general, model of multi-agent systems, and showed how the histories traced out in the execution of such a system could be used as the semantic foundation for a family of both linear and branching time temporal belief logics. He then gave examples of how these logics could be used in the specification and verification of protocols for cooperative action. 2. 7 Communication Formalisms for representing communication in agent theory have tended to be based on speech act theory, as originated by Austin (1962), and further developed by Searle (1969) and others (Cohen & Perrault, 1979; Cohen & Levesque, 1990a). Briefly, the key axiom of speech act theory is that communicative utterances arc actions, in just the sense that physical actions arc. They are performed by a speaker with the intention of bringing about a desired change in the world: typically, the speaker intends to bring about some particular mental state in a listener. Speech acts may fail in the same way that physical actions may fail: a listener generally has control over her mental state, and cannot be guaranteed to react in the way that the speaker intends, Much work in speech act theory has been devoted to classifying the various different types of speech acts. Perhaps the two most widely recognised categories of speech acts are representatives (o f which informing is the paradigm example), and directives (of which requesting is the paradigm example). Although not directly based on work in speech acts (and arguably more to do with architectures than theories), we shall here mention work on agent communication languages (Genesereth & Ketchpel, 1994). The best known work on agent communication languages is that by the ARPA knowledge sharing effort (Patil et al,, 1992). This work has been largely devoted to developing two related languages: the knowledge query and manipulation language (KQML) and the knowledge

M. WOOLDRIDGE AND NICHOLAS JENNINGS 126 An all-embracing agent theory is some time off, and yet signficant steps have been taken towards it. In the following subsections, we briefly review some of this work. 2.6. I Moore-knowledge and action Moore was in many ways a pioneer of the use of logics for capturing aspects of agency (Moore, 1990). His main concern was the study of knowledge pre-conditions for actions-the question of what an agent needs to know in order to be able to perform some action. He formalised a model of ability in a logic containing a modality for knowledge, and a dynamic logic-like apparatus for modelling action (cf. Hare!, 1984). This formalism allowed for the possibility of an agent having incomplete information about how to achieve some goal, and performing actions in order to find out how to achieve it. Critiques of the formalism (and attempts to improve on it) may be found in Morgenstern (1987) and Lesperance (1989). 2.6.2 Cohen and Levesque-intention One of the best-known and most influential contributions to the area of agent theory is due to Cohen and Levesque (1990a). Their formalism was originally used to develop a theory of intention (as in "I intend to ... "), which the authors required as a pre-requisite for a theory of speech acts (Cohen & Levesque, 1990b). However, the logic has subsequently proved to be so useful for reasoning about agents that it has been used in an analysis of conflict and cooperation in multi agent dialogue (Galliers, 1988a,b), as well as several studies in the theoretical foundations of cooperative problem solving (Levesque ct al., 1990; Jennings, 1992; Castelfranchi, 1990; Castel franchi et al., 1992). Here, we shall review its use in developing a theory of intention. Following Bratman (1990), Cohen and Levesque identify seven properties that must be satisfied by a reasonable theory of intention: 1. Intentions pose problems for agents, who need to determine ways of achieving them. 2. Intentions provide a "filter" for adopting other intentions, which must not conflict. 3. Agents track the success of their intentions, and arc inclined to try again if their attempts fail. 4. Agents believe their intentions are possible. 5. Agents do not believe they will not bring about their intentions. 6. Under certain circumstances, agents believe they will bring about their intentions. 7. Agents need not intend ail the expected side effects of their intentions. Given these criteria, Cohen and Levesque adopt a two-tiered approach to the problem of formalising intention. First, they construct a logic of rational agency, "being careful to sort out the relationships among the basic modal operators" (Cohen & Levesque, 1990a, p. 221). Over this framework, they introduce a number of derived constructs, which constitute a "partial theory of rational action" (Cohen & Levesque, 1990a, p. 221); intention is one of these constructs. The first major derived construct is the persistent goal. An agent has a persistent goal of rp iff: l. It has a goal that q; eventually becomes true, and believes that rp is not currently true. 2. Before it drops the goal cp, one of the following conditions must hold: i the agent believes cp has been satisfied; or ii the agent believes cp will never be satisfied. It is a small step from persistent goals to a first definition of intention, as in '•intending to act'': an agent intends to do action a iff it has a persistent goal to have brought about a state wherein it believed it was about to do (1, and then did a. Cohen and Levesque go on to show how such a definition meets manyof Bratman'scritcria for a theory of intention (outlined above). A critique of Cohen and Levesque's theory of intention may be found in Singh (1992). 2.6.3 Rao and Georgeff-belief, desire, intention architectures As we observed earlier, there is no clear consensus in either the Al or philosophy communities about precisely which combination of information and pro-attitudes are best suited to characteris ing rational agents. In the work of Cohen and Levesque, described above, just two basic attitudes

M. WOOLDRIDGE AND NICHOLAS JENNINGS 124 from belief: it seems reasonable that one could believe something that is false, but one would hesitate to say that one could know something false. Knowledge is thus often defined as true belief; i knows cp if i believes <p and <pis true. So defined, knowledge satisfies T. Axiom 4 is called the positive introspection axiom. Introspection is the process of examining one's own beliefs, and is discussed in detail in (Konolige, 1986a, Chapter 5). The positive introspection axiom says that an agent is aware of what it knows. Similarly, axiom 5 is the negative introspective axiom, which says that an agent is aware of what it doesn't know. Positive and negative introspection together imply an agent has perfect knowledge about what it does and doesn't know (cf. (Konolige, 1986a, Equation (5.11), p. 79)). Whether or not the two types of introspection are appropriate properties for knowledge/belief is the subject of sonie debate. However, it is generally accepted that positive introspection is a less demanding property than negative introspection, and is thus a more reasonable property for resource bounded reasoners. Given the comments above, the axioms KTD45 are often chosen as a logic of (idealised) knowledge, and KD45 as a logic of (idealised) belief. 2.4 Alternatives to the possible worlds model As a result of the difficulties with logical omniscience, many researchers have attempted to develop alternative formalisms for representing belief. Some of these are attempts to adapt the basic possible worlds model; others represent significant departures from it. In the subsections that follow, we examine some of these attempts. 2.4.1 Levesque-belief and awareness In a 1984 paper, Levesque proposed a solution to the logical omniscience problem that involves making a distinction between explicit and implicit belief (Levesque, 1984). Crudely, the idea is that an agent has a relatively small set of explicit beliefs, and a very much larger (infinite) set of implicit beliefs, which includes the logical consequences of the explicit beliefs. To fonnalise this idea, Levesque developed a logic with two operators; one each for implicit and explicit belief. The semantics of the explicit belief operator were given in terms of a weakened possible worlds semantics, by borrowing some ideas from situation semantics (Barwise & Perry, 1983; Devlin, 1991). The semantics of the implicit belief operator were given in terms of a standard possible worlds approach. A number of objections have been raised to Levesque's model (Reichgelt, 1989b. p. 135): first, it does not allow quantification-this drawback has been rectified by Lakemeycr (1991); second, it docs not seem to allow for nested beliefs; third, the notion of a situation, which underlies Levesque's logic is, if anything, more mysterious than the notion of a world in possible worlds; and fourth, under certain circumstances, Levesque's proposal still makes unrealistic predictions about agent's reasoning capabilities. In an effort to recover from this last negative result, Fagin and Halpern have developed a "logic of general awareness" based on a similar idea to Levesque's but with a very much simpler semantics (Fagin & Hapern, 1985). However, this proposal has itself been criticised by some (Konolige, 1986b). 2.4.2 Konolige-the deduction model A more radical approach to modelling resource bounded believers was proposed by Konolige (Konolige, 1986a). His deduction model of belief is, in essence, a direct attempt to model the "beliefs" of :.ymbolic Al systems. Konolige observed that a typical knowledge-based system has two key components: a database of symbolically represented "beliefs" (which may take the form of rules. frames, semantic nets, or, more generally, formulae in some logical language), and some logically incomplete inference mechanism. Konolige modelled such systems in terms of deduction structures. A deduction structure is a pair d = (ii, p), where ~ is a base set of formulae in some logical language, and pis a set of inference rules (which may be logically incomplete), representing the agent's reasoning mechanism. To simplify the formalism, Konolige assumed that an agent

Intelligent agents: theory and practice 123 complete axiomatisation of normal modal logic. Similarly, the second property will appear as a rule of inference in any axiomisation of normal modal logic; it is generally called the necessitation rule. These two properties turn out to be the most problematic features of normal modal logics when they are used as logics of knowledge/belief (this point will be examined later). The most intriguing properties of normal modal logics follow from the properties of the accessibility relation, R, in models. To illustrate these properties, consider the following axiom schema: □<p =-<p. It turns out that this axiom is characteristic of the class of models with a reflexive accessibility relation. (By characteristic, we mean that it is true in all and only those models in the class.) There are a host of axioms which correspond to certain properties of R: the study of the way that properties of R correspond to axioms is called correspondence theory. For our present purposes, we identify just four axioms: the axiom called T (which corresponds to a reflexive accessibility relation); D (serial accessibility relation); 4 (transitive accessibility relation); and 5 (euclidean accessibility relation): T □q; => q; D □,p ~ ◊'P 4 □q; => □□q; 5 ◊'P ~ □◊'PThe results of correspondence theory make it straightforward to derive completeness results for a range of simple normal modal logics. These results provide a useful point of comparison for normal modal logics, and account in a large part for the popularity of this style of semantics. To use the logic developed above as an epistemic logic, the formula □q: is read as: "it is known that rp". The worlds in the model are interpreted as epistemic alternatives, the accessibility relation defines what the alternatives arc from any given world. The logic defined above deals with the knowledge of a single agent. To deal with multi-agent knowledge, one adds to a model structure an indexed set of accessibility rehitions, one for each agent. The language is then extended by replacing the single modal operator "O" by an indexed set of unary modal operators { K1}, where i E {1 , ... , n }. The formula K;r:r is read: •'i knows that cp". Each operator K, is given exactly the same properties as·'□". The next step is to consider how well normal modal logic serves as a logic of knowledge/belief. Consider first the necessitation rule and axiom K, since any normal modal system is committed to these. The necessitation rule tells us that an agent knows all valid formulae. Amongst other things, this means an agent knows all propositional tautologies. Since there is an infinite number of these, an agent will have an infinite number of items of knowledge: immediately, one is faced with a counter-intuitive property of the knowledge operator. Now consider the axiom K, which says that an agent's knowledge is closed under implication. Together with the necessitation rule, this axiom implies that an agent's knowledge is closed under logical consequence: an agent believes all the logical consequences of its beliefs. This also seems counter intuitive. For example, suppose, like every good logician, our agent knows Pcano's axioms. Now Fermat's last theorem follows from Pean o's axioms-but it took the combined efforts of some of the best minds over the past century to prove it. Yet if our agent's beliefs are closed under logical consequence, then our agent must know it. So consequential closure, implied by necessitation and the K axiom, seems an overstrong property for resource bounded reasoners. These two problems-that of knowing all valid formulae, and that of knowledge/belief being closed under logical consequence-together constitute the famous logical omniscience problem. It has been widely argued that this problem makes the possible worlds model unsuitable for representing resource bounded believers-and any real system is resource bounded. 2.3.1 Axioms for knowledge and belief We now consider the appropriateness of the axioms D. T, 4, and 5 for logics of knowledge/ belief. The axiom D says that an agent's beliefs are non-contradictory; it can be re-written as: K,cp => -.K, ..,<p, which is read: '•if i knows rp, then i doesn't know -irp'". This axiom seems a reasonable property of knowledge/belief. The axiom Tis often called the knowledge axiom, since it says that what is known is true. It is usually accepted as the axiom that distinguishes knowledge

M. WOOLDRIDGE AND NICHOLAS JENNINGS 122 developed by Kripke (1963).3 Hintikka's insight was to see that an agent's beliefs could be characterised as a set of possible worlds, in the following way. Consider an agent playing a card game such as poker.4 In this game, the more one knows about the cards possessed by one's opponents, the better one is able to play. And yet complete knowledge of an opponent's cards is generally impossible (if one excludes cheating). The ability to play poker well thus depends, at least in part, on the ability to deduce what cards are held by an opponent, given the limited information available. Now suppose our agent possessed the ace of spades. Assuming the agent's sensory equipment was functioning normally, it would be rational of her to believe that she possessed this card. Now suppose she were to try to deduce what cards were held by her opponents. This could be done by first calculating all the various different ways that the cards in the pack could possibly have been distributed among the various players. (This is not being proposed as an actual card playing strategy, but for illustration!) For argument's sake, suppose that each possible configuration is described on a separate piece of paper. Once the process is complete, our agent can then begin to systematically eliminate from this large pile of paper all those configurations which are not possible, given what she knows. For example, any configuration in which she did not possess the ace of spades could be rejected immediately as impossible. Call each piece of paper remaining after this process a world. Each world represents one state of affairs considered possible, given what she knows. Hintikka coined the term epistemic alternatives to describe the worlds possible given one's beliefs. Something true in all our agent's epistemic alternatives could be said to be believed by the agent. For example, it will be true in all our agent's epistemic alternatives that she has the ace of spades. On a first reading, this seems a peculiarly roundabout way of characterising belief. but it has two advantages. First, it remains neutral on the subject of the cognitive structure of agents. It certainly doesn't posit any internalised collection of possible worlds. It is just a convenient way of characterising belief. Second, the mathematical theory associated with the formalisation of possible worlds is extremely appealing (see below). The next step is to show how possible worlds may be incorporated into the semantic framework of a logic. Epistemic logics arc usually formulated as normal modal logics using the semantics developed by Kripke (1963). Before moving on to explicitly epistemic logics, we consider a simple normal modal logic. This logic is essentially classical propositional logic, extended by the addition of two operators:'·□" (necessarily), and"." (possibly). Let Prop= {p, q, .. . } be a countable set of atomic propositions. Then the syntax of the logic is defined by the following rules: (i) if p e Prop then p isa formula; (ii) if sigmpahi, are formulae, then so are -sigmaand i:p V 1.j J; and (iii) if rp i~ a formula then so arc □qi and ◊ff. The operators "-," (not) and "V" (o r) have their standard meanings. The remaining connectives of classical propositional logic can be defined as abbreviations in the usual way. The formula □q, is read: "necessarily cp" and the formula ◊ff is read: "possibly cp". The semantics of the modal connectives arc given by introducing an accessibility relation into models for the language. This relation defines what worlds are considered accessible from every other world. The formula □qi is then true if q; is true in every world accessible from the current world; ◊rp is true if rp is true in at least one world accessible from the current world. The two modal operators are duals of each other, in the sense that the universal and existential quantifiers of first-order logic arc duals: It would thus have been possible to take either one as primitive, and introduce the other as a derived operator. The two basic properties of this logic arc as follows. First, the following axiom = = schema is valid: □(q-, 1.j J) (Oq, -=O1.JJ). This axiom is called K, in honour of Kripkc. The second property is as follows: if <pis valid, then □q; is valid. Now, since K is valid, it will be a theorem of any 3In Hintikka's original work. he used a technique based on "model sets which is equivalent to Kripke's formalism, though less elegant. See Hughes and Cresswell (1968, pp. 351-352) for a comparison and discussion of the two techniques. 4This example was adapted from Halpern (1987).

Intelligent agents: theory and practice 121 logic fail here? The problem is that the intentional notions-such as belief and desire-are referentially opaque, in that they set up opaque contexts, in which the standard substitution rules of first-order logic do not apply. In classical (propositional or first-order) logic, the denotation, or semantic value, of an expression is dependent solely on the denotations of its sub-expressions. For example, the denotation of the propositional logic formulap /\ q is a function of the truth-values of p and q. The operators of classical logic are thus said to be truth functional. In contrast, intentional notions such as belief are not truth functional. It is surely not the case that the truth value of the sentence: Janine believes p (5) is dependent solely on the truth value of p 2 So substituting equivalents into opaque contexts is not going to preserve meaning. This is what is meant by referential opacity. Clearly, classical logics are not suitable in their standard form for reasoning about intentional notions: alternative formalisms are required. The number of basic techniques used for alternative formalisms is quite small. Recall, from the discussion above, that there arc two problems to be addressed in developing a logical formalism for intentional notions: a syntatic one, and a semantic one. It follows that any formalism can be characterised in terms of two independent attributes: its language of formulation, and semantic model (Konolige, 1986a, p. 83). There are two fundamental approaches to the syntactic problem. The first is to use a modal language, which contains non-truth-functional modal operators, which arc applied to formulae. An alternative approach involves the use of a meta-language: a many-sorted first-order language containing terms that denote formulae of some other object-language. Intentional notions can be represented using a meta-language predicate, and given whatever axiomatisation is deemed appropriate. Both of these approaches have their advantages and disadavantagcs, and will be discussed in the sequel. As with the syntactic problem, there arc two basic approaches to the semantic problem. The first, best-known, and probably most widely used approach is to adopt a possible worlds semantics, where an agent's beliefs, knowledge, goals, and so on, arc characterised as a set of so-called possible worlds, with and accessibility relation holding between them. Possible worlds semantics have an associated correspondence theory which makes them an attractive mathematical tool to work with (Chellas, 1980). However, they also have many associated difficulties, notably the well known logical omniscience problem, which implies that agents are perfect reasoners (we discuss this problem in more detail below). A number of variations on the possible-worlds theme have been proposed, in an attempt to retain the correspondence theory, but without logical omnis cience. The commonest alternative to the possible worlds model for belief is to use a sentential, or interpreted symbolic structures approach. In this scheme, beliefs are viewed as symbolic formulae explicitly represented in a data structure associated with an agent. An agent then believes sig i m f asigmias present in its belief data structure. Despite its simplicity, the sentential model works well under certain circumstances (Konolige, 1986a). In the subsections that follow, we discuss various approaches in some more detail. We begin with a close look at the basic possible world:-, model for logics of knowledge (episremic logics) and logics of belief (doxastic logics). 2.3 Possible worlds semantics The possible worlds model for logics of knowledge and belief was originally proposed by Hintikka (1962), and is now most commonly formulated in a normal modal logic using the techniques 2Note, however, that the sentence (5) is itself a proposition, in that its denotation is the value true or false.

M. WOOLDRIDGE AND NICHOLAS JENNINGS 120 Being an intentional system seems to be a necessary condition for agenthood. but is it a sufficient condition? In his Master's thesis, Shardlow trawled through the literature of cognitive science and its component disciplines in an attempt to find a unifying concept that underlies the notion of agenthood. He was forced to the following conclusion: "Perhaps there is something more to an agent than its capacity for beliefs and desires, but whatever that thing is, it admits no unified account within cognitive science." (Shardlow, 1990) So, an agent is a system that is most conveniently described by the intentional stance; one whose simplest consistent description requires the intentional stance. Before proceeding, it is worth considering exactly which attitudes are appropriate for representing agents. For the purposes of this survey, the two most important categories are information attitudes and pro-attitudes: desire intention obligation commitment intention belief obligation information attitudes pro-attitudes l { commitment knowledge choice Thus information attitudes are related to the information that an agent has about the world it occupies, whereas pro-attitudes are those that in some way guide the agent's actions. Precisely which combination of attitudes is most appropriate to characterise an agent is, as we shall sec later, an issue of some debate. However, it seems reasonable to suggest that an agent must be represented in terms of at least one information attitude, and at least one pro-attitude. Note that pro- and information attitudes are closely !inked, as a rational agent will make choices and form intentions, etc., on the basis of the information it has about the world. Much work in agent theory is concerned with sorting out exactly what the relationship between the different attitudes is. The next step is to investigate methods for representing and reasoning about intentional notions. 2.2 Representing intentional notions Suppose one wishes to reason about intentional notions in a logical framework. Consider the following statement (after Genesereth & Nilsson, 1987, pp. 210-211): Janine believes Cronos is the father of Zeus. (1) A naive attempt to translate (1) into first-order logic might result in the following: Bel(Janine, Father(Zeus, Cronos)) (2) Unfortunately, this naive translation does not work, for two reasons. The first is syntactic: the second argument to the Bel predicate is a formula of first-order logic, and is not, therefore, a term. So (2) is not a well-formed formula of classical first-order logic. The second problem is semantic, and is potentially more serious. The constants Zeus and Jupiter, by any reasonable interpretation, denote the same individual: the supreme deity of the classical world. It is therefore acceptable to write, in first-order logic: (Zeus= Jupiter). (3) Given (2) and (3), the standard rules of first-order logic would allow the derivation of the following: Bel(Janine, Father(Jupiter, Cronos)) (4) But intuition rejects this derivation as invalid: believing that the father of Zeus is Cronos is not the same as believing that the father of Jupiter is Cronos. So what is the problem? Why does first-order

Intelligent agents: theory and practice 119 These statements make use of a folk psychology, by which human behaviour is predicted and explained through the attribution of attitudes, such as believing and wanting (as in the above examples), hoping, fearing and so on. This folk psychology is well established: most people reading the above statements would say they found their meaning entirely clear, and would not give them a second glance. The attitudes employed in such folk psychological descriptions are called the intentional notions. The philosopher Daniel Dennett has coined the term intentional system to describe entities "whose behaviour can be predicted by the method of attributing belief, desires and rational acumen" (Dennett, 1987, p. 49). Dennett identifies different "grades" of intentional system: "A first-order intentional system has beliefs and desires (etc.) but no beliefs and desires (and no doubt other intentional states) about beliefs and desires .... A second-order intentional system is more sophisticated; it has beliefs and desires (and no doubt other intentional states) about beliefs and desires (and other intentional state)-both those of others and its own" (Dennett, 1987, p. 243) One can carry on this hierarchy of intentionality as far as required. An obvious question is whether it is legitimate or useful to attribute beliefs, desires, and so on, to artificial agents. Isn't this just anthropomorphism? Mc Carthy, among others, has argued that there are occasions when the intentional stance is appropriate: "To ascribe beliefs, free will, intentions, consciousness, abilities, or wants to a machine is legitimate when such an ascription expresses the same information about the machine that it expresses about a person. It is useful when the ascription helps us understand the structure of the machine, its past or future behaviour, or how to repair or improve it. It is perhaps never logically required even for humans, but expressing reasonably briefly what is actually known about the state of the machine in a particular situation may require mental qualities or qualities isomorphic to them. Theories of belief, knowledge and wanting can be constructed for machines in a simpler setting than for humans, and later applied to humans. Ascription of mental qualities is most straightforward for machines of known structure such as thermostats and computer operating systems, but is most useful when applied to entities whose structure is incompletely known." (Mc Carthy, 1978) (quoted in (Shoham, 1990)) What objects can be described by the intentional stance? As it turns out, more or less anything can. In his doctoral thesis, Seel showed that even very simple, automata-like objects can be consistently ascribed intentional descriptions (Seel 1989); similar work by Rosenschein and Kaelbling (albeit with a different motivation), arrived at a similar conclusion (Rosenschein & Kaelbling, 1986). For example, consider a light switch: "It is perfectly coherent to treat a light switch as a (very cooperative) agent with the capability of transmitting current at will, who invariably transmits current when it believes that we want it transmitted and not otherwise; flicking the switch is simply our way of communicating our desires.'' (Shoham, 1990, p. 6) And yet most adults would find such a description absurd-perhaps even infantile. Why is this? The answer seems to be that while the intentional stance description is perfectly consistent with the observed behaviour of a light switch, and is internally consistent, .. it does not buy us anything, since we essentially understand the mechanism sufficiently to have a simpler, mechanistic description of its behaviour." (Shoham, 1990, p. 6) Put crudely, the more we know about a system, the less we need to rely on animistic, intentional explanations of its behaviour. However, with very complex systems, even if a complete, accurate picture of the system's architecture and working is available, a mechanistic, design stance explanation of its behaviour may not be practicable. Consider a computer. Although we might have a complete technical description of a computer available, it is hardly practicable to appeal to such a description when explaining why a menu appears when we click a mouse on an icon. In such situations, it may be more appropriate to adopt an intentional stance description, if that description is consistent, and simpler than the alternatives. The intentional notions are thus abstraction tools, which provide us with a convenient and familiar way of describing, explaining, and predicting the behaviour of complex systems.

M. WOOLDRIDGE AND NICHOLAS JENNINGS 118 convenience, we identify three key issues, and structure our survey around these (cf. Seel, 1989, p.1 ): • Agent theories are essentially specifications. Agent theorists address such questions as: How are we to conceptualise agents? What properties should agents have, and how are we to formally represent and reason about these properties? • Agent architectures represent the move from specification to implementation. Those working in the area of agent architectures address such questions as: How are we to construct computer systems that satisfy the properties specified by agent theorists? What software and/or hardware structures are appropriate? What is an appropriated separation of concerns? • Agent languages are programming languages that may embody the various principles proposed by theorists. Those working in the area of agent languages address such questions as: How are we to program agents? What are the right primitives for this task? How are we to effectively compile or execute agent programs? As we pointed out above, the distinctions between these three areas are occasionally unclear. The issue of agent theories is discussed in the section 2. In section 3, we discuss architectures, and in section 4, we discuss agent languages. A brief discussion of applications appears in section 5, and some concluding remarks appear in section 6. Each of the three major sections closes with a discussion, in which we give a brief critical review of current work and open problems, and a section pointing the reader to further relevant reading. Finally, some notes on the scope and aims of the article. First, it is important to realise that we are writing very much from the point of view of AI, and the material we have chosen to review clearly reflects this bias. Secondly, the article is not a intended as a review of Distributed AI, although the material we discuss arguably falls under this banner. We have deliberately avoided discussing what might be called the macro aspects of agent technology (i.e., those issues relating to the agent society, rather than the individual (Gasser, 1991), as these issues are reviewed more thoroughly elsewhere (see Bond and Gasser, 1988, pp. 1-56, and Chaibdraa et al., 1992). Thirdly, we wish to reiterate that agent technology is, at the time of writing, one of the most active areas of research in AI and computer science generally. Thus, work on agent theories, architectures, and languages is very much ongoing. In particular, many ofthe fundamental problems associated with agent technology can by no means be regarded as solved. This article therefore represents only a snapshot of past and current work in the field, along with some tentative comments on open problems and suggestions for future work areas. Our hope is that the article will introduce the reader to some of the different ways that agency is treated in D(AI), and in particular to current thinking on the theory and practice of such agents. 2 Agent theories In the preceding section, we gave an informal overview of the notion of agency. In this section, we turn our attention to the theory of such agents, and in particular, to formal theories. We regard an agent theory as a specification for an agent; agent theorists develop formalisms for representing the properties of agents, and using these formalisms, try to develop theories that capture desirable properties of agents. Our starting point is the notion of an agent as an entity 'which appears to be the subject of beliefs, desires, etc.' (Seel, 1989, p. 1). The philosopher Dennett has coined the term intentional system to denote such systems. 2. 1 Agents as intentional systems When explaining human activity, it is often useful to make statements such as the following: Janine took her umbrella because she believed it was going to rain. Michael worked hard because he wanted to possess a Ph D.

Intelligent agents: theory and practice 117 A simple way of conceptualising an agent is thus as a kind of UNIX-like software process, that exhibits the properties listed above. This weak notion of agency has found currency with a surprisingly wide range of researchers. For example, in mainstream computer science, the notion of an agent as a self-contained, concurrently executing software process, that encapsulates some state and is able to communicate with other agents via message passing, is seen as a natural development of the object-based concurrent programming paradigm (Agha, 1986; Agha ct al., 1993). This weak notion of agency is also that used in the emerging discipline of agent-based software engineering: "[Agents} communicate with their peers by exchanging messages in an expressive agent communication Language. While agents can be as simple as subroutines, typically they are larger entities with some sort of persistent control." (Gcncscrcth & Kctchpcl, 1994, p.48) A softbot (software robot) is a kind of agent: "A softbot is an agent that interacts with a software environment by issuing commands and interpreting the environments feedback. A softbot's effectors arc commands (e.g. Unix shell commands such as mv or compress) meant to change the external environments state. A softbot's sensors are commands (e.g. pwd or ls in Unix) meant to provide . . information." (Etzioni et al., 1994, p.10) 1.1.2 A stronger notion of agency For some researchers-particularly those working in AI- the term "agent" has a stronger and more specific meaning than that sketched out above. These researchers generally mean an agent to be a computer system that, in addition to having the properties identified above, is either conceptualised or implemented using concepts that arc more usually applied to humans. For example, it is quite common in AI to characterise an agent using mentalistic notions, such as knowledge, belief, intention, and obligation (Shoham, 1993). Some AI researchers have gone further, and considered emotional agents (Bates et al., 1992a; Bates, 1994). (Lest the reader suppose that this is just pointless anthropomorphism, it should be noted that there are good arguments in favour of designing and building agents in terms of human-like mental states-sec section 2.) Another way of giving agents human-like attributes is to represent them visually, perhaps by using a cartoon-like graphical icon or an animated face (Maes, 1994a, p. 36)-for obvious reasons, such agents are of particular importance to those interested in human-computer interfaces. 1.1.3 Other attributes of agency Various other attributes arc sometimes discussed in the context of agency. For example: • mobility is the ability of an agent to move around an electronic network (White, 1994); • veracity is the assumption that an agent will not knowingly communicate false information (Galliers, 1988b, pp. 159-164); • benevolence is the assumption that agents do not have conflicting goals, and that every agent will therefore always try to do what is asked of it (Rosenschein and Genesereth, 1985, p. 91); and • rationality is (crudely) the assumption that an agent will act in order to achieve its goals, and will not act in such a way as to prevent its goals being achieved-at least insofar as its beliefs permit (Galliers, 1988b, pp. 49-54). (A discussion of some of these notions is given below; various other attributes of agency are formally defined in (Goodwin, 1993).) 1.2 The structure of this article Now that we have at least a preliminary understanding of what an agent is, we can embark on a more detailed look at their properties, and how we might go about constructing them. For

M. WOOLDRIDGE AND NICHOLAS JENNINGS 116 control has long been a research domain in distributed artificial intelligence (DAI) (Steeb et al., 1988); various types of information manager, that filter and obtain information on behalf of their users, have been prototyped (Maes, 1994a); and systems such as those that appear in the third scenario are discussed in (Mc Gregor, 1992; Levy ct al., 1994). The key computer-based com ponents that appear in each of the above scenarios are known as agents. It is interesting to note that one way of defining Al is by saying that it is the subfield of computer science which aims to construct agents that exhibit aspectsof intelligent behaviour. The notion of an "agent" is thus central to AI. It is perhaps surprising, therefore, that until the mid to late 1980s, researchers from mainstream AI gave relatively little consideration to the issues surrounding agent synthesis. Since then, however, there has been an intense flowering of interest in the subject: agents are now widely discussed by researchers in mainstream computer science, as well as those working in data communications and concurrent systems research, robotics, and user interface design. A British national daily paper recently predicted that: "Agent-hased computing (ABC) is likely to be the next significant breakthrough in software development.·· (Sargent, 1992) Moreover, the UK-based consultancy firm Ovum has predicted that the agent technology industry would be worth some US$3.5 billion worldwide by the year 2000 (Houlder, 1994). Researchers from both industry and academia arc thus taking agent technology seriously: our aim in this paper is to survey what we perceive to be the most important issues in the design and construction of intelligent agents, of the type that might ultimate appear in applications such as those suggested by the fictional scenarios ahove. We begin our article, in the following sub-section, with a discussion on the subject of exactly what an agent is. I. I What is an agent? Carl Hewitt recently remarked1 that the question what is an agent? is embarrassing for the agent based computing community in just the same way that the question what is intelligence? is embarrassing for the mainstream AI community. The problem is that although the term is widely used, by many people working in closely related areas, it defies attempts to produce a single universally accepted definition. This need not necessarily be a problem: after all, if many people are successfully developing interesting and useful applications, then it hardly matters that they do not agree on potentially trivial terminological details. However, there is also the danger that unless the issue is discussed, "agent" might become a "noise'" term, subject to both abuse and misuse, to the potential confusion of the research community. It is for this reason that we briefly consider the question. We distinguish two general usages of the term "agent": the first is weak, and relatively uncontentious; the second is stronger, and potentially more contentious. I. I. I A Weak Notion of Agency Perhaps the most general way in which the term agent is used is to denote a hardware or (more usually) software-based computer system that enjoys the following properties: • autonomy: agents operate without the direct intervention of humans or others, and have some kind of control over their actions and internal state (Castelfranchi, 1995); • social ability: agents interact with other agents (and possibly humans) via some kind of agent communication language (Genesereth & Ketchpel, 1994); • reactivity: agents perceive their environment (which may be the physical world, a user via a graphical user interface, a collection of other agents, the Internet, or perhaps all of these combined), and respond in a timely fashion to changes that occur in it; • pro-activeness: agents do not simply act in response to their environment, they are able to exhibit goal-directed behaviour by taking the initiative. 1A t the Thirteenth International Workshop on Distributed Al.

The Knowledge Engineering Review, Vol. 10:2, 1995, 115-152 Intelligent agents: theory and practice MICHAEL WOOLDRIDGE1 and NICHOLAS R. JENNINGS2 1 Department of Computing Manchester Metropolitan University Chester Street, Manchester MI 5GD, UK (M. Wooldridge@doc.mmu.ac.uk) Department of Electronic F.ngineering, Queen Mary & Westfield College, Mile End Road, London El 4N.S, UK ( N. R.Jennings@gm w.ac.uk) Abstract The concept of an agent has become important in both artificial intelligence (AI) and mainstream computer science. Our aim in this paper is to point the reader at what we perceive to be the most important theoretical and practical issues associated with the design and construction of intelligent agents. For convenience, we divide these issues into three areas (though as the reader will see, the divisions arc at times somewhat arbitrary). Agent theory is concerned with the question of what an agent is, and the use of mathematical formalisms for representing and reasoning about the properties of agents. Agent architectures can he thought of as software engineering models of agents; researchers in this area are primarily concerned with the problem of designing software or hardware systems that will satisfy the properties specified by agent theorists. Finally, agent languages are software systems for programming and experimenting with agents; these languages may embody principles proposed by theorists. The paper is not intended to serve as a tutorial introduction to all the issues mentioned; we hope instead simply to identify the most important issues, and point to work that elaborates on them. The article includes a short review of current and potential applications of agent technology. 1 Introduction We begin our article with descriptions of three events that occur sometime in the future: I. The key air-traffic control systems in the country of Ruritania suddenly fail, due to freak weather conditions. Fortunately, computerised air-traffic control systems in neighbouring countries negotiate between themselves to track and deal with all affected flights, and the potentially disastrous situation passes without major incident. 2. Upon logging in to your computer, you are presented with a list of email messages, sorted into order of importance by your personal digital assistant (PDA). You are then presented with a similar list of news articles; the assistant draws your attention to one particular article, which describes hitherto unknown work that is very close to your own. After an electronic discussion with a number of other PD As, your PDA has already obtained a relevant technical report for you from an FTPsite, in the anticipation that it will be of interest. 3. You are editing a file, when your PDA requests your attention: an email message has arrived, that containsnotification about a paper you sent to an important conference, and the PDA correctly predicted that you would want to sec it as soon as possible. The paper has been accepted, and without prompting, the PDA begins to look into travel arrangements, by consulting a number of databases and other networked information sources. A short time later, you are presented with a summary of the cheapest and most convenient travel options. We shall not claim that computer systems of the sophistication indicated in these scenarios are just around the corner, but serious academic research is underway into similar applications: air-traffic

M. WOOLDRIDGE AND NICHOLAS JENNINGS 152 Weerasooriya, D, Rao, A and Ramamohanarao, K, 1995. "Design of a concurrent agent-oriented language" In: M Wooldridge and NR Jennings (eds.) Intelligent Agents: Theories, Architectures, and Languages (LNAJ Volume 890), pp 386----402, Springer-Verlag. Weihmayer, Rand Velthuijsen, H, 1994. "Application of distributed AI and cooperative problem solving to telecommunications" In: J Liebowitz and D Prcreau (eds.) Al Approaches to Telecommunications and Network Management, IOS Press. Werner, E, 1988. "Toward a theory of communication and cooperation for multiagent planning" In: MY Var di (ed.) Proceedings of the Second Conference on Theoretical Aspects of Reasoning About Knowledge, pp 129-144, Morgan Kaufmann. Werner, E, 1989. "Cooperating agents: A unified theory of communication and social structure" lq: L Gasser and M Huhns (eds.) Distributed Artificial Intelligence Volume 11, pp 3---36, Pitman. Werner, E, 1990. "Wh.at·can agents do together: A semantics of cooperative ability" In: Proceedings of the Ninth European Conference on Artificial Intelligence (ECAJ-90), pp 694-701, Stockholm, Sweden. Werner, E, 1991. "A unified view of information, intention and ability" In: Y Demazeau and JP Miiller (eds.) Decentralized Al 2-Proceedings of the Second European Workshop on Modelling Autonomous Agents and Multi-Agent Worlds (MAAMA W-90), pp 109-126, Elsevier. White, JE, 1994. "Telescript technology: The foundation for the electronic marketplace'', White paper, General Magic, Inc., 2465 Latham Street, Mountain View, CA 94040. Wilkins, D, 1988. Practical Planning: Extending the Classical Al Planning Paradigm, Morgan Kaufmann. Wittig, T (ed.) 1992. ARCHO N: An Architecture for Multi-Agent Systems, Ellis Horwood. Wood, S, 1993. Planning and Decision Making in Dynamic Domains, Ellis Horwood. Wooldridge, M, 1992. The Logical Modelling of Computational Multi-Agent Systems, Ph D thesis, Depart ment of Computation, UMIST, Manchester, UK. (Also available as Technical Report MMU-DOC-94-01, Department of Computing, Manchester Metropolitan University, Chester Street, Manchester, UK.) Wooldridge, M, 1994. ·"Coherent social action" In: Proceedings of the Ele J.'enth European Conference on Artificial Intelligence (ECAI-94), pp 279-283, Amsterdam, The Netherlands. Wooldridge, M, 1995. "This is MYWORLD: The logic of an agent-oriented testbed for DAI" In: M Wooldridge and NR Jennings (eds.) Intelligent Agents: Theories, Architectures, and Languages (LNAI Volume 890), pp 160-178, Springer-Verlag. Wooldridge, M and Fisher M, 1992. "A first-order branching time logic of multi-agent systems" In: Proceedings of the Tenth European Conference on Artificial Intelligence (ECAl-92), pp 234-238, Vienna, Austria. Wooldridge, Mand Fisher M, 1994. "A decision procedure for a temporal belief logic" In: OM Gabbay and HJ Ohlbach. (eds.) Temporal Logic-Proceedings of the First International Conference (LNAI Volume 827), pp 317-331, Springer-Verlag. Wooldridge, M and Jennings NR, 1994. "Formalizing the cooperative problem solving process" In: Proceedings of the Thirteenth International Workshop on Distributed Artificial Intelligence (IWDAJ-94), pp 403---417, Lake Quinalt, WA. Yonczawa, A (ed.) 1990. ABCL-An Object-Oriented Concurrent System, MIT Press.

Intelligent agents: theory and practice 151 Sargent, P, 1992. "Back to school for a brand new ABC" In: The Guardian, 12 March, p 28. Schoppe N;, MJ, 1987. "Universal plans for reactive robots in unpredictable environments" In: Proceedings of the Tenth International Joint Conference on Artificial Intelligence ( /JCAl-87), pp 1039-1046, Milan, Italy. Schwuttke, UM and Quan, AG, 1993. '·Enhancing performance of cooperating agents in real-time diagnostic systems" In: Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence (JJCAl93), pp 332-337, Chambfay, France. Searle, JR, 1969. Speech Acts: An Essay in the Philosophy of Language, Cambridge University Press. Seel, N, 1989. Agent Theories and Architectures, Ph D thesis, Surrey University, Guildford, UK. Segerbcrg, K, 1989. "Bringing it about" Journal of Philosophical Logic 18 327-347. Shardlow, N, 1990. "Action and agency in cognitive science", Master's thesis, Department of Psychology, University of Manchester, Oxford Road, Manchester M13 9PL, UK. Shoham, Y, 1988. Reasoning About Change: Time and Causation from the Standpoint of A rtificial Intelligence, MIT Press. Shoham, Y, 1989. "Time for action: on the relation between time, knowledge and action" In: Proceedings of the Eleventh /nternational Joint Conferenceon Artificial Intelligence (l JCAl-89), pp 954-959, Detroit, Ml. Shoham, Y, 1990. "Agent-oriented programming", Technical Report STAN-CS-1335-90, Computer Science Department, Stanford University, Stanford, CA 94305. Shoham, Y, 1993. "Agent-oriented programming" Artificial Intelligence 60 (1) 51-92. Singh, MP, l990a. '·Group intentions" In: Proceedings of the Tenth International Workshop on Distributed Artificial Intelligence (IWDAI-90). Singh, MP. 1990b. "Towards a theory of situated know-how" In: Proceedings of the Ninth European Conference on Artificial Intelligence (ECAJ-90), pp 604-609, Stockholm, Sweden. Singh, MP, 1991a. "Group ability and structure" In: Y Dcmazeau and JP MUiler (eds.) Decentralized Al 2Proceedings of the Second European Workshop on Modelling Autonomous Agents and Multi-Agent Worlds (MAAMAW-90), pp 127-146, Elsevier. Singh, MP, 1991b '·Towards a formal theory of communication for multi-agent systems" In: Proceedings of rhe Twelfth International Joint Conference on Artificial Intelligence (l JCAI-91), pp 69-74, Sydney, Australia. Singh, MP, 1992. "A critical examination of the Cohen-Levesque theory of intention" In: Proceedings of the Tenth European Conference on Artificial Intelligence (ECAl-92), pp 364-368. Vienna, Austria. Singh, MP, 1994. Multiagent Systems: A Theoretical Framework for Intentions, Know-How, and Communi cations (LNAI Volume 799), Springer-Verlag. Singh, MP and Asher, NM, 1991. "Towards a formal theory of intentions" In: Logics in Al-Proceedings of the European Workshop JELIA-90 (LNA! Volume 478), pp 472-486, Springer-Verlag. Smith, RG, 1980. A Framework for Distributed Problem Solving, UMI Research Press. Steeb, R, Cammarata S, Hayes-Roth FA, Thorndyke PW and Wesson RB, 1988. "Distributed intelligence for air fleet control" In: AH Bond and L Gasser (eds.) Readings in Distributed Artificial Intelligence, pp 90101, Morgan Kaufmann. Steels, L, 1990. "Cooperation between distributed agents through self organization" In: Y Demazeau and JP MUiler (eds.) Decentralized Al-Proceedings of the First European Workshop on Modelling Autonomous Agents in Multi-Agent Worlds (MAAMA W-89), pp 175-196, Elsevier. Thomas, SR, 1993. PLACA, an Agent Oriented Programming Language, Ph D thesis, Computer Science Department, Stanford University, Stanford, CA 94305. (Available as technical report STAN-CS-931487). Thoma~, SR, Shoham Y, Schwartz A and Kraus S, 1991. "Preliminary thoughts on an agent description language" International Journal of Intelligent Systems 6 497-508. Thomason, R, 1980. "A note on syntactical treatments of modality" Synthese 44 391-395. Turner, R, 1990. Truth and Modality for Knowledge Representation, Pitman. Varga, LZ, Jennings, NR and Cockburn, D, 1994. "Integrating intelligent systems into a cooperating community for electricity distribution management" International Journal of Expert Systems with Appli cations 1 (4) 563-579. Vere, Sand Bickmore, T, 1990. "A basic agent" Computational Intelligence 6 41-60. Voorhees, EM, 1994. •'Software agents for information retrieval" In: 0 Etzioni (ed.) Software Agents Papers from the 1994 Spring Symposium (Technical Report SS-94-03), pp 126--129, AAAI Press. Wainer, J, 1994. "Yet another semantics of goals and goal priorities" In: Proceedings of the Eleventh European Conference on Artificial Intelligence (ECIA-94), pp 269-273, Amsterdam, The Netherlands. Wavish, P, 1992. "Exploiting emergent behaviour in multi-agent systems" In: E Werner and Y Demazeau (eds.) Decentralized Al 3-Proceedings of the Third European Workshop on Modelling Autonomous Agents and Multi-Agent Worlds (MAAMA W-91), pp 297-310, Elsevier. Wavish, P and Graham, M, 1995. "Role, skills, and behaviour: a situated action approach to organising systems of interacting agents" In: M Wooldridge and NR Jennings (eds.) Intelligent Agents: Theories, Archi1ectures, and Languages (LNAI Volume 890), pp 371-385, Springer-Verlag.

M. WOOLDRIDGE AND N[CHOLAS JENNINGS 150 MUiler, JP and Pischel, M, 1994. "Modelling interacting agents in dynamic environments .. In: Proceedings of the Eleventh European Conference on Artificial Intelligence (ECAI-94), pp 709-713, Amsterdam, The Netherlands. Mi.i\ler, JP, Pischcl, M and Thiel, M, 1995. "Modelling reactive behaviour in vertically layered agent architectures" In: M Wooldridge and NR Jennings (eds.) Intelligent Agents: Theories, Architectures, and Languages (LNAI Volume 890), pp 261-276, Springer-Verlag. Newell, A and Simon, HA, 1976. "Computer science as empirical enquiry" Communications of the ACM 19 113-126. Nilsson, NJ, 1992. "Towards agent programs with circuit semantics", Technical Report STAN-CS-92-1412, Computer Science Department, Stanford University, Stanford, CA 94305. Norman, TJ and Long, D, 1995. "Goal creation in motivated agents" In: M Wooldridge and NR Jennings (eds.) Intelligent Agents: Theories, Architectures, and Languages (LNAI Volume 890), pp 277-290, Springer-Verlag. Papazoglou, MP, Laufman, SC and Sel\is, TK, 1992. "An organizational framework for cooperating intelligent information systems" Journal of Intelligent and Cooperative Information Systems 1 (1) 169-202. Parunak, HVD, 1995. "Applications of distributed artificial intelligence in industry'" In: GMP O'Hare and NR Jennings (eds.) Foundations of Distributed Al, John Wiley. Patil, RS, Fikes, RE, Patel-Schneider, PF, Mc Kay, D, Finin, T, Gruber, T and Neches, R, 1992. "The DARPA knowledge sharing effort: Progress report" In: C Rich, W Swartout and B Nebel (eds.) Proceedings of Knowledge Representation and Reasoning (KR&R-92), pp 777-788. Perlis, D, 1985. "Languages with self reference I: Foundations" Artificial Intelligence 25 301-322. Perlis, D, 1988. "Languages with self reference II: Knowledge, belief, and modality" Artificial Intelligence 34 179-212. Perloff, M, 1991. "STIT and the language of agency" Synthese 86 379-408. Poggi, A, 1995. "DAISY: An object-oriented system for distributed artificial intelligence" In· M Wooldridge and NRJennings (eds.) Intelligent Agents: Theories, Architectures, and Languages (LNA!Volume890). pp 341-354, Springer-Verlag. Pollack, ME and Ringuette, M, 1990. "Introducing the Tileworid: Experimentally evaluating agent architec tures" In: Proceedings of the Eighth National Conference on Artificial Intelligence (AAAl-90), pp 183-189, Boston, MA. Rao, AS and Georgeff, MP, 1991a. '"Asymmetry thesis and side-effect problems in linear time and branching time intention logics'" In: Proceedings oft he Twelfth International Joint Conference on Artificial Intelligence (IJCAl-91), pp 498-504, Sydney, Australia. Rao, AS and Georgeff, MP, 1991b. "Modeling rational agents within a BDI-architccture·• In: R Fikes and E Sandewal! (eds.) Proceedings of Knowledge Representation and Reasoning (KR&R-9/), pp 473-484, Morgan Kaufmann. Rao, AS and Georgeff, MP, 1992a. "An abstract architecture for rational agents"' In: C Rich, W Swartout and B Nebel (eds.) Proceedings of Knowledge Representation and Reasoning (KR&R-92), pp 439-449. Rao, AS and Georgeff, MP, 1992b. "Social plans: Preliminary report" In: E Werner and Y Demazeau (eds.) Decentralized Al 3-Proceedings of the Third European Workshop on Modelling Autonomous Agents and Multi-Agent Worlds (MAAMA W-9I), pp 57-76, Elsevier. Rao, AS and Georgeff, MP, 1993. "A model-theoretic approach to the verification of situated reasoning systems" In: Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence (IJCAI93), pp 318-324, Chambery, France. Reichgclt, H, 1989a "A comparison of first-order and modal logics of time" In: P Jackson, H Reichge!t and F van Harmclen (eds.) Logic Based Knowledge Representation, pp 143-176, MIT Press. Reichgelt. H, 198%. "Logics for reasoning about knowledge and belief" Knowledge Engineering Review4 (2) 119-139. Rosenschein, JS and Genesereth, MR, 1985 "Deals among rational agents·• In: Proceedings of the Ninth International Joint Conference on Anificial Intelligence (Il CAl-85), pp 91-99, Los Angeles, CA Rosenschein, S, 1985. "Formal theories of knowledge in AI and robotics'' New Generation Computing, pp 345-357. Rosenschein, S and Kae\bling, LP, 1986. '·Toe synthesis of digital machines with provable epistemic properties" In: JY Halpern (ed.) Proceedings of the 1986 Conference on Theoretical Aspects of Reasoning About Knowledge, pp 83-98, Morgan Kaufmann. Russell, SJ and Wefald, E, 1991 Do the Right Thing-Studies in Limited Rationality. MIT Press. Sacerdoti. E, 1974. "Planning in a hierarchy of abstraction spaces" Artificial Intelligence 5 115-135. Sacerdoti, E, 1975. '"The non-linear nature of plans'' In: Proceedings of the Fourth International Joint Conference on Artificial Intelligence ( Il CAl-75), pp 206--214. Stanford, CA. Sadek. MD, 1992. '·A study in the logic of intention" In: C Rich, W Swartout and B Nebel (eds.) Proceedings of Knowledge Representation and Reasoning (KR&R-92). pp 462-473.

Intelligent agents: theory and practice 149 Kiss, G and Reichgelt, H, 1992. "Towards a semantics of desires" In: E Werner and Y Demazeau (eds.) Decentralized Al 3-Proceedings of the Third European Workshop on Modelling Autonomous Agents and Multi-Agent Worlds (MAAMAW-91), pp 115-128, Elsevier. Konolige, K, 1982. "A first-order formalization of knowledge and action for a multi-agent planning system" In: JE Hayes, D Michie and Y Pao (eds.) Machine Intelligence JO, pp 41-72, Ellis Horwood. Konolige, K, 1986a. A Deduction Model of Belief, Pitman. Konolige, K, 1986b. "What awareness isn't: A sentential view of implicit and explicit belief (position paper)" In: JY Halpern (ed.) Proceedings of the 1986 Conference on Theoretical Aspects of Reasoning About Knowledge, pp 241-250, Morgan Kaufmann. Konolige, Kand Pollack, ME, 1993. "A representationalist theory of intention" In: Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence (IJCAJ-93), pp 390--395, Chambefy, France. Kraus, Sand Lehmann, D (1988) "Knowledge, belief and time" Theoretical Computer Science 58 155-174. Kripke, S, 1963. "Semantical analysis of modal logic" Zeitschrift f Ur Mathematische Logik und Grundlagen der Mathematik 9 67-96. Lakemeyer, G, 1991. "A computationally attractive first-order logic of belief" In: JELIA-90: Proceedings of the European Workshop on Logics in AI (LNAI Volume 478), pp 333-347, Springer-Verlag. Lesperance, Y, 1989. "A formal account of self knowledge and action" In: Proceedings of the Eleventh Imernational Joint Conference on Artificial Intelligence (IJCAI-89), pp 868-874, Detroit, Ml. Levesque, HJ, 1984. "A logic of implicit and explicit belief" In: Proceedings oft he Fourth National Conference on Artificial Intelligence (AAAI-84), pp 198-202, Austin, TX. Levesque, HJ, Cohen, PR and Nunes, JHT, 1990. "On acting together" In: Proceedings of the Eighth National Conference on Artificial Intelligence (AAAJ-90), pp 94-99, Boston, MA. Levy, A Y, Sagiv, Y and Srivastava, D, 1994. "Towards efficient information gathering agents" In: 0 Etzioni (ed.) Software Agents-Papers from the 1994 Spring Symposium (Technical Report SS-94-03), pp 64-70, AAAI Press. Mack, D, 1994. "A new formal model of belief" In: Proceedings of the Eleventh European Conference on Artificial Intelligence (ECAI-94), pp 573-577, Amsterdam, The Netherlands. Maes, P, 1989. "The dynamics of action selection" In: Proceedings of the Eleventh International Joint Conference on Artificial Intelligence (IJCAI-89), pp 991-997, Detroit, MI. Maes, P (ed.) 1990a. Designing Autonomous Agents, MIT Press. Maes, P, 1990b. "Situated agents can have goals" In: P Maes (ed.) Designing Autonomous Agents, pp49-70, MIT Press. Maes, P, 1991. "The agent network architecture (ANA)" SIGA RT Bulletin 2 (4) 115-120. Maes, P, 1994a. "Agents that reduce work and information overload" Communications of the ACM 37 (7) 3140. Macs, P, 1994b. "Social interface agents: Acquiring competence by learning from users and other agents" In: 0 Etzioni (ed.) Software Agents-Papers from the J 994 Spring Symposium (Technical Report SS-94-03), pp 71-78, AAA! Press. Mc Cabe, FG and Clark, KL, 1995. "April-agent process interaction language" In: M Wooldridge and NR Jennings (eds.) Intelligent Agents: Theories, Architectures, and Languages (LNAI Volume 890), pp 324340, Springer-Verlag. Mc Carthy, J, 1978. "Ascribing mental qualities to machines." Technical report, Stanford University Al Lab., Stanford, CA 94305. Mc Gregor, SL, 1992. "Prescient agents" In: D Coleman (ed.) Proceedings of Groupware-92, pp 228-230. Montague, R, 1963. "Syntactical treatments of modality, with corollaries on reflexion principles and finite axiomatizations" Acta Philosophica Fennica 16153-167. Moore, RC, 1990. '"A formal theory of knowledge and action" In: JF Allen, J Hendler and A Tate (eds.) Readings in Planning, pp 480-519, Morgan Kaufmann. Morgenstern, L, 1987. "Knowledge preconditions for actions and plans" In: Proceedings of the Tenth International Joint Conference on Artificial Intelligence (JJCAI-87), pp 867-874, Milan, Italy. Mori, K, Torikoshi, H, Nakai, Kand Masuda, T, 1988. "Computer control system for iron and steel plants" Hitachi Review 37 (4) 251-258. Morley, RE and Schelberg, C, 1993. "An analysis of a plant-specific dynamic scheduler'· In: Proceedings of the NSF Workshop on Dynamic Scheduling, Cocoa Beach, Florida. Mukhopadhyay, U, Stephens, Land Huhns, M, 1986. "An intelligent system for document retrieval in distributed office environments'' Journal of the American Society for Information Science 37 123-135. Millier, JP, 1994. "A conceptual model for agent interaction" In: SM Deen (ed.) Proceedings of the Second International Working Conference on Cooperating Knowledge Based Systems (CKBS-94), pp 213-234, DAKE Centre, University of Keele, UK.

M. WOOLDRIDGE AND NICHOLAS JENNINGS 148 Halpern, JY, 1986. "Reasoning about knowledge: An overview" In: JY Halpern (ed.) Proceedings oft he 1986 Conference on Theoretical Aspects of Reasoning About Knowledge, pp 1-18, Morgan Kaufmann. Halpern, JY, 1987. "Using reasoning about knowledge to analyze distributed systems" Annual Review of Computer Science 2 37--68. Halpern, JY and Moses, Y, 1992. "A guide to completeness and complexity for modal logics of knowledge and belief' Artificial Intelligence 54 319-379. Halpern, JY and Vardi, MY, 1989. "The complexity of reasoning about knowledge and time. I. Lower bounds" Journal of Computer and System Sciences 38 195-237. Hare!, D, 1984. "Dynamic logic" In: D Gabbay and F Guenther (eds.) Handbook of Philosophical Logic Volume ll-Extensions of Classical Logic, pp 497-604, Reidel. Haugeneder, H, 1994. IMAGINE final project report. Haugeneder, Hand Steiner, D, 1994. "A multi-agent approach to cooperation in urban traffic" In: SM Deen (ed.) Proceedings of the 1993 Workshop on Cooperating Knowledge Based Systems (CKBS-93), pp 83-98, DAKE Centre, University of Keele, UK. Haugeneder, H, Steiner, D and Mc Cabe, FG, 1994. "IMAGINE: A framework for building multi-agent systems" In: SM Deen (ed.) Proceedings of the 1994 Jnternational Working Conference on Cooperating Knowledge Based Systems (CKBS-94), pp 31--64, DAKE Centre, University of Keele, UK. Hayes-Roth, B, 1990. "Architectural foundations for real-time performance in intelligent agents" The Journal of Real-Time Systems 2 99-125. Hendler, J (ed.) 1992. Artificial intelligence Planning: Proceedings of the First International Conference, Morgan Kaufmann. Henz, M, Smolka, G and Wuertz, J, 1993. "Oz-a programming language for multi-agent systems" ln: Proceedings of the Thirteenth international Joint Conference on Artificial Intelligence ( IJCAI-93), pp 404409, Chamb Cry, France. Hewitt, C, 1977. "Viewing control structures as patterns of passing messages" Artificial intelligence 8 (3) 323364. Hintikka, J, 1962. Knowledge and Belief, Cornell University Press. Houlder, V, 1994. "Special agents" In: Financial Times, 15 August, p 12. Huang, J, Jennings, NR and Fox. J, 1995. "An ag-ent architecture for distributed medical care" In: M Wooldridge and NR Jennings (eds.) Intelligent Agents: Theories, Architectures, and Languages (LNA! Volume 890), pp 219-232, Springer-Verlag. Hughes, GE and Cresswell, MJ, 1968. lntroduction to Modal Logic, Methuen. Huhns, MN, Jacobs, N, Ksiezyk, T, Shen, WM, Singh, MP and Cannata, PE, 1992. "Integrating enterprise information models in Carnot" In: Proceedings of the International Conference on Intelligent and Cooperative information Systems, pp 32-42, Rotterdam, The Netherlands. Israel, DJ, 1993. "The role(s) of logic in artificial intelligence" In: DM Gabbay, CJ Hogger and JA Robinson (eds.) Handbook of Logic in Artificial Intelligence and Logic Programming, pp 1-29, Oxford University Press. Jennings, NR, 1992. "On being responsible" In: E Werner and Y Demazeau (eds.) Decentralized Al] Proceedings of the Third European Workshop on Modelling Autonomous Agents and Multi-Agent Worlds (MAAMA W-91), pp 93-102, Elsevier. Jennings. NR, 1993a. "Commitments and conventions: The foundation of coordination in multi-agent systems" Knowledge Engineering Review 8 (3) 223-250. Jennings, NR, 1993b. "Specification and implementation of a belief desire joint-intention architecture for collaborative problem solving" Journal of intelligent and Cooperative Information Systems 2 (3) 289-318. Jennings, NR, 1995. "Controlling cooperative problem solving in industrial multi-agent systems using joint intentions" Artificial Intelligence 14 (2) (to appear). Jennings, NR. Varga, LZ, Aarnts, RP, Fuchs, J and Skarek, P, 1993. "Transforming standalone expert systems into a community of cooperating agents'' International Journal of Engineering Applications of Artificial Intelligence 6 (4 ) 317-331. Kaelbling, LP, 1986. "An architecture for intelligent reactive systems" In: MP George ff and AL Lansky (eds.) Reasoning About Actions and Plans-Proceedingofthe 1986 Workshop, pp 395-410, Morgan Kaufmann. Kaelbling, LP, 1991. ·'A situated automata approach to the design of embedded agents" SIG ART Bulletin 2 (4) 85-88. Kaelbling, LP and Roscnschein, SJ, 1990. "Action and planning in embedded agents" In: P Maes (ed.) Designing Autonomous Agents, pp 35-48, MIT Press. Kinny, D, Ljungberg, M, Rao, AS, Sonenberg, E, Tidhar, G and Werner, E, 1992. "Planned team activity" In: C Castelfranchi and E Werner (eds.) Artificial Social Systems-Selected Papers from the Fourth European Workshop on Modelling Autonomous Agents and Multi-Agent Worlds, MAAAMA W-92 (LNA! Volume 830), pp 226-256, Springer-Verlag.

Intelligent agents: theory and practice 147 Fikes, RE and Nilsson. N, 1971. "STRIPS: A new approach to the application of theorem proving to problem solving'· Artificial Intelligence 5 (2) 189-208. Firby, JA. 1987. "An investigation into reactive planning in complex domains"' In: Proceedings of the Tenth International Joint Conference on Artificial Intelligence (IJCA/-87), pp 202-206, Milan, Italy. Fischer, K, Kuhn. N, MUiler, HJ, MUiler, JP and Pischel, M, 1993. ·'Sophisticated and distributed: The transportation domain" In: Proceedings of the Fifth European Workshop on Modelling Autonomous Agents and Multi-Agent Worlds (MAAMA W-93), Neuchatel, Switzerland. Fisher, M, 1994. "A survey of Concurrcnt Metate M-the language and its applications"" In: DM Gabbay and HJ Ohlbach (eds.) Temporal Logic-Proceedings of the First International Conference (LNAJ Volume 827). pp 480-505, Springer-Verlag. Fisher, M. 1995. "Representing and executing agent-based systems·· In: M Wooldridge and NR Jennings (eds.) Intelligent Agents: Theories, Architectures, and Languages (LNAJ Volume 890), pp 307-323, Springer-Verlag. Fisher, M and Wooldridge, M, 1993. "Specifying and verifying distributed intelligent systems" In: M Filgueiras and L Damas (eds.) Progress in Artificial Intelligence-Sixth Portuguese Conference on Anificial Intelligence (LNAI Volume 727), pp 13-28. Springer-Verlag. Galliers. JR, 1988a. "A strategic framework for multi-agent cooperative dialogue" In: Proceedings of the Eighth European Conference on Artificial Intelligence (ECAl-88), pp 415-420, Munich, Germany. Galliers, JR, 1988b. A Theoretical Framework for Computer Models of Cooperative Dialogue, Acknowledg ing Multi-Agent Conflict. Ph D thesis, Open University, UK. Gasser, L, 1991. "Social conceptions of knowledge and action: DAI foundations and open systems semantics" Artificial Intelligence 47 107-138. Gasser, L, Braganza, C and Hermann, N, 1987. "MACE: A flexible testbed for distributed AI research" In: M Huhns (ed.) Distributed Artificial Intelligence, pp 119-152, Pitman. Gasser, Land Briot, JP, 1992. "Object-based concurrent programming and DAI" In: Distributed Artificial Intelligence: Theory and Praxis. pp 81-108, Kluwer Academic. Geissler, C and Konolige. K, 1986. "A resolution method for quantified modal logics of knowledge and helief'· In: JY Halpern (ed.) Proceedings of the 1986 Conference on Theoretical Aspects of Reasoning About Knowledge, pp 309-324, Morgan Kaufmann. Genescreth, MR and Ketchpe L SP, 1994. "Software agents" Communications of the ACM 37 (7) 48-53. Genesereth. MR and Nils~on, N, 1987. Logical Foundations of Artificial Intelligence, Morgan Kaufmann. Gcorgeff, MP, 1987. "Planning" Annual Review of Computer Science 2 359-400. Georgeff, MP and lngrand, FF, 1989. "Decision-making in an embedded reasoning system" In: Proceedings of the Eleventh International Joint Conference on Artificial Intelligence ( IJCAl-89), pp 972-978, Detroit, ML Gcorgcff. MP and Lansky, AL (eds.) 1986. Reasoning About Actions & Plans-Proceedings of the 1986 Workshop, Morgan Kaufmann. Georgeff, MP and Lansky, AL, 1987. "Reactive reasoning and planning In: Proceedings of the Sixth National Conference on Artificial Intelligence (AAAJ-87), pp 677-682, Seattle, WA. Ginsberg, M. 1993. Essentials of Artificial Intelligence, Morgan Kaufmann. Gmytrasiewicz, P and Durfee. EH, 1993. "Elements of a utilitarian theory of knowledge and action" In: Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence (IJCA/-93), pp 396-- 402, Chamb Cry. France. Goldblatt, R. 1987. Logics of Time and Computation, Centre for the Study of Language and lnformation Lccturc Notes Series. (Distributed by Chicago University Press.) Goldman, RP and Lang. RR, 1991 "Intentions in time"', Technical Report TUTR 93-101, Tulane University. Goodwin, R. 1993. "Formalizing properties of agents", Technical Report CMU-CS-93-159, School of Computer Science. Carnegie-Mellon University, Pittsburgh, PA. Greif. I, 1994. "Desktop agent~ in group-enabled products" Communications of the ACM 37 (7) 100-105. Grosz, BJ and Sidner, CL, 1990. ''Plans for discourse" In: PR Cohen, J Morgan and ME Pollack (eds.) Intentions in Communication. pp 417-444, MIT Press. Gruber. TR, 1991. "The role of common ontology in achieving sharable, reusable knowledge bases'· In: R Fikes and E Sandcwall (eds.) Proceedings of Knowledge Representation and Reasoning (KR&R-91), Morgan Kaufmann. Guha, RV and Lcnat, DB, 1994. "Enabling agents to work together" Communications of the ACM 37 (7) 127142. Haas, A, 1986. "A syntactic theory of belief and knowledge" Artificial Intelligence 28 (3) 245-292. Haddadi, A. 1994. "A hybrid architecture for multi-agent systems" In: SM Deen (ed.) Proceedings of the 1993 Workshop 011 Cooperating Knowledge Based Systems (CKBS-93), pp 13-26. DAKE Centre, University of Keele, UK.

M. WOOLDRIDGE AND NICHOLAS JENNINGS 146 Chaib-draa, B, Moulin, B, Mandiau, Rand Millot, P, 1992. "Trends in distributed artificial intelligence·· Artificial Intelligence Review 6 35-66. Chang, E, 1987. "Participant systems" In: M Huhns (ed.) Distributed Artificial Intelligence, pp 311-340, Pitman. Chapman, D, 1987. "Planning for conjunctive goals" Artificial Intelligence 32 333--378. Chapman, D and Agre, P, 1986. "Abstract reasoning as emergent from concrete activity" In: MP Georgeff and AL Lansky (eds.) Reasoning About Actions & Plans~Proceedings of the 1986 Workshop pp 411-424, Morgan Kaufmann. Chel\as, B, 1980. Modal Logic: An Introduction, Cambridge University Press. Chu, D, 1993. "l.C. PROLOG JI: A language for implementing multi-agent systems" In: SM Deen (ed.) Proceedings of the 1992 Workshop on Cooperating Knowledge Based Systems (CKBS-92), pp 61-74, DAKE Centre, University of Keele, UK. Cohen, PR, Greenberg ML, Hart OM and Howe AE. 1989. "Trial by fire: Understanding the design requirements for agents in complex environments" Al Magazine 10 (3) 32-48. Cohen, PR and Levesque, HJ, 1990a. "'Intention is choice with commitment" Artificial Intelligence 42 213261. Cohen, PR and Levesque, HJ, 1990h. ·'Rational interaction as the basis for communication" In: PR Cohen, J Morgan and ME Pollack (eds.) Intentions in Communication, pp 221-256. MIT Press. Cohen, PR and Perrault, CR, 1979. ·'Elements of a plan based theory of speech acts" Cognitive Science 3177212. Connah, D and Wavish. P, 1990. '·An experiment in cooperation" In: Y Demazeau and J-P Miiller (eds.) Decentralized AI-Proceedings of the First European Workshop on Modelling Autonomous Agents in Multi-Agent Worlds (MAAMA W-89), pp 197-214, Elsevier. Cutkosky, MR, Engelmorc, RS, Fikes. RE, Gcnesereth. MR, Gruber, T, Mark, WS, Tenenbaum, JM and Weber, JC, 1993. "PACT: An experiment in integrating concurrent engineering systems'' IEEE Computer 26 (l) 28-37. Davies, NJ, 1993. Truth, Modality, and Action, Ph D thesis, Department of Computer Science, University of Essex, Colchester, UK. Dean, TL and Wellman, MP. 1991. Planning and Control, Morgan Kaufmann. Dennett, DC, 1978. Brainstorm~, MIT Press. Dennett, DC. 1987. The Intentional Stance, MIT Press. des Rivieres. J and Levesque. HJ, 1986. ''The consistency of ~yntactical treatments of knowledge" In: JY Halpern ( ed,) Proceedings of the 1986 Conference on Theoretical Aspects of Reasoning About Knowledge, pp 115-130, Morgan Kaufmann. Devlin, K, 1991. Logic and Information, Cambridge University Press. Dongha, P, 1995 "Toward a formal model of commitment for rc~ource-bounded agent~" In: M Wooldridge and NR Jennings (eds.) Intelligent Agents: Theories, Architectures, and Languages ( LNAJ Volume 890), pp 86--101. Springer-Verlag. Down~, J and Rcichgelt, II, 1991. "Integrating classical and rcm.:tive planning within an architecture for autonomous agents·· In: J Hertzberg (ed,) European Worhhop on Planning (LNAI Volume 522), pp 13-- 26. Doyle. J, Shoham, Y and Wellman, MP, 1991. ''A logic of relative desire" In: ZW Ras and M Zemankova (cd~.) Methodologies for Intelligent Systems-Sixth International Symposium, ISMIS-91 (LNAI Volume 542). Springer-Verlag. Emerson EA, 1990. "Temporal and modal logic" In: J van Leeuwen (ed.) Handhookof Throretical Computer Science. pp 996--1072, Elsevier. Emerson, EA and Halpern, JY, 1986. "'Sometimes' and 'not never' revisited: on branching time versus linear time temporal logic" Journal of the ACM 33 (1) 151-178. Etzioni, 0. Lesh, N and Segal, R, 1994. ·'Building soft bots for UNIX'' In: 0 Etzioni (ed.) Software Agents Paper~ from the 1994 Spring Symposium (Technical Report SS-94-03), pp 9-16, AAAI Press. Fagin, Rand Halpern, JY, 1985. "Belief, awarenes~. and limited reasoning" In: Proceedings of the Ninth International Joint Conference on Artificial Intelligence (IJCAl-85), pp 480--490. Los Angeles, CA. Fagin. R, Halpern, JY and Yard!, MY, 1992. "What can machines know? on the properties of knowledge in distributed system~·· Journal of rhe A CM 39 (2) 328--376. Ferguson IA. 1992a. Touring Machines: An Architecture for Dynamic, Rational, Mobile Agents, Ph D thesis, Clare Hall, University of Cambridge, UK. (Also available as Technical Report No. 273, University of Cambridge Computer Laboratory.) Ferguson. IA, 1992h. ·Towar<ls an architecture for adaptive, rational, mobile agents·• In: E Werner and Y Demazeau (eds.) Decentralized Al 3-Proceedings of the Third European Workshop on Modelling Autonomous Agents and Multi-Agent Worlds (MAAMA W-91), pp 249-262, Elsevier.

Intelligent agents: theory and practice 145 Agre, P and Chapman, D, 1987. "PENGI: An implementation of a theory of activity" In: Proceedings of the Sixth National Conference on Artificial Intelligence (AAAI-87), pp 268-272, Seattle, WA. Allen, JF, 1984. "Towards a general theory of action and time" Artificial Intelligence 23 (2) 123--154. Allen, JF, Hendler, J and Tate, A (eds.), 1990. Readings in Planning. Morgan Kaufmann. Allen, JF, Kautz, H, Pelavin, Rand Tenenberg, J, 1991. Reasoning About Plans. Morgan Kaufmann. Ambros-Ingerson, J and Steel, S, 1988. "Integrating planning, execution and monitoring" In: Proceedings of the Seventh National Conference on Artificial Intelligence (AAAI-88), pp 83-88, St. Paul, MN. Austin, JL, 1962. How to Do Things With Words. Oxford University Press. Ay!ett, Rand Eustace, D, 1994. "Multiple cooperating robots-combining planning and behaviours'' In: SM Deen (ed) Proceedings of the 1993 Workshop on Cooperating Knowledge Based Systemv (CKBS-93), pp 3-- 11. DAKE Centre, University of Keele, UK. Haecker, RM (ed.) 1993. Readings in Groupware and Computer-Supported Cooperative Work. Morgan Kaufmann. Barringer, H, Fisher, M, Gabbay, D, Gough, G and Owens, R, 1989. "Metate M: A framework for programming in temporal logic" In: REX Workshop on Stepwise Refinement of Distributed Systems: Models, Formalisms, Correctness (LNCS Volume 430) pp 94-129. Springer-Verlag. Barwise, J and Perry, J, 1983. Situations and Attitudes, MIT Press. Bates, J, 1994. "The role of emotion in believable agents" Communications of the ACM 37 (7) 122-125. Bates, J, Bryan Loyall, A and Scott Reilly, W, 1992a. "An architecture for action, emotion, and social behaviour". Technical Report CMU-CS-92-144, School of Computer Science, Carnegie-Mellon Univer sity, Pittsburgh, PA. Bates, J, Bryan Loyall, A and Scott Reilly, W, 19926. "Integrating reactivity, goals, and emotion in a broad agent". Technical Report CMU-CS-92-142, School of Computer Science, Carnegie-Mellon University, Pittsburgh, PA. Bell, J, 1995. "Changing attitudes". In: M Wooldridge and NR Jennings (eds.) Intelligent Agents: Theories, Architectures, and Languages (LNAI Volume 890), pp 40--55, Springer-Verlag. Belnap, N, 1991. ''Backwards and forwards in the modal logic of agency'' Philosophy and Phenomenological Research LI (4) 777-807. Belnap, N and Perloff, M, 1988. "Seeing to it that: a canonical form for agentives" Theoria 54175-199. Bond, AH and Gasser, L (eds.) 1988. Readings in Distribwed Artificial Intelligence, Morgan Kaufmann. Bratman, ME, 1987. Intentions, Plans, and Practical Reason, Harvard University Press. Bratman, ME, 1990. "What is intention?" In: PR Cohen, JL Morgan and ME Pollack (eds.) Intentions in Communication, pp 15-32, MIT Press. Bratman, ME, Israel, DJ and Pollack, ME, 1988. "Plans and resource-bounded practical reasoning'' Computational Intelligence 4 349-355. Brooks, RA, 1986. "A robust layered control system for a mobile robot" IEEE Journal of Robotics and Aulomation 2 (1) 14-23. Brooks, RA, 1990. "Elephants don't play chess" In: P Maes (ed.) Designing Aulonomous Agents, pp 3-15, MIT Press. Brooks, RA, 1991a. "Intelligence without reason" In: Proceedings of the Twelflh International Joint Conference on Artificial Intelligence (JJCAl-91), pp 569-595, Sydney, Australia. Brooks, RA, 19916. "Intelligence without representation" Artificial Intelligence 47 139-159. Burmeister, Band Sundermeyer, K. 1992. ''Cooperative problem solving guided by intentions and percep tion" In: E Werner and Y Demazeau (eds) Decentralized Al 3-Proceedings of the Third European Workshop on Modelling Autonomous Agents and Multi-Agent Worlds (MAAMAW-91), pp 77-92, Elsevier. Bussman, Sand Demazeau, Y, 1994. "An agent model combining reactive and cognitive capabilities" In: Proceedings of the IEEE International Conference on Intelligent Robots and Systems (IROS-94), Munich, Germany. Caste!franchi, C, 1990. "Social power" In: Y Demazeau and J-P MU!ler(eds.) Decentralized Al-Proceedings of the First European Workshop on Modelling Autonomous Agents in Multi-Agent Worlds (MAAMA W8Y), pp 49--62, Elsevier. Castelfranchi, C, 1995. "Guarantees for autonomy in cognitive agent architecture" In: M Wooldridge and NR Jennings (eds.) Intelligent Agents. Theories, Architectures, and Languages (LNAI Volume 890), pp 56--70, Springer-Verlag. Castelfranchi, C, Miceli, Mand Cesta, A, 1992. '·Dcpcndcncc relations among autonomous agents" In: E Werner and 'Y Dema:ceau (eds.) Decentralized Al 3-Proceedings of the Third European Workshop on Mode Uing Awonomous Agents and Multi-Agent Worlds (MAAMA W-Yl), pp 215-231, Elsevier. Catach, L, 1988. "Norma! multimodal logics"' In: Proceedings of the Seventh National Conference on Anificial Intelligence (AAAJ-88), pp 491-495. St. Paul, MN.

M. WOOLDRIDGE AND NICHOLAS JENNINGS 144 specified articles from a range of document repositories (Voorhees, 1994). Another important system in this area is called Carnot (Huhns et al., 1992), which allows pre-existing and hetero geneous database systems to work together to answer queries that are outside the scope of any of the individual databases. 5.4 Believable agents There is ohvious potential for marrying agent technology with that of the cinema, computer games, and virtual reality. The Oz project6 was initiated to develop: .. artistically interesting, highly interactive, simulated worlds . . to give users the experience of living in (not merely watching) dramatically rich worlds that include moderately competent, emotional agents" (Batesetal., 1992b,p. 1) To construct such simulated worlds, one must first develop believable agents: agents that "provide the illusion oflife, thus permitting the audience's suspension of disbelief" (Bates, 1994, p. 122). A key component of such agents is emotion: agents should not be represented in a computer game or animated film as the flat, featureless characters that appear in current computer games. They need to show emotions; to act and react in a way that resonates in tune with our empathy and understanding of human behaviour. The Oz group have investigated various architectures for emotion (Bates et al., 1992a), and have developed at least one prototype implementation of their ideas (Bates, 1994). 6 Concluding remarks This paper has reviewed the main concepts and issues associated with the theory and practice of intelligent agents. It has drawn together a very wide range of material, and has hopefully provided an insight into what an agent is, how the notion of an agent can be formalised, how appropriate agent architectures can be designed and implemented, how agents can be programmed, and the types of applications for which agent-based solutions have been proposed. The subject matter of this review is important because it is increasingly felt, both within academia and industry, that intel!igent agents will be a key technology as computing systems become ever more distributed, interconnected, and open. In such environments, the ability of agents to autonomously plan and pursue their actions and goals, to cooperate, coordinate, and negotiate with others, and to respond flexibly and intelligently to dynamic and unpredictable situations will lead to significant improve ments in the quality and sophistication of the software systems that can be conceived and implemented, and the application areas and problems which can be addressed. Acknowledgements Much of this paper was adapted from the first author's 1992 Ph D thesis (Wooldridge, 1992), and as such this work was supported by the UK Science and Engineering Research Council (now the EPSRC). We arc grateful to those people who read and commented on earlier drafts of this article, and in particular to the participants of the 1994 workshop on agent theories, architectures, and languages for their encouragement, enthusiasm, and helpful feedback. Finally, we would like to thank the referees of this paper for their perceptive and helpful comments. References Adorni, G and Poggi, A, 1993. "An object-oriented language for distributed artificial intelligence" Inter national Journal of Man-Machine Studies 38 435-453. Agha, G, 1986. ACTORS: A Model of Concurrent Computation in Distributed Systems. MIT Press. Agha, G, Wegner, P and Yonezawa, A (eds.), 1993. Research Directions in Concurrent Objec1-0riented Programming. MIT Pre:.s. 6 Not to be confused with the Oz programming language (Henz ct al., 1993).

Intelligent agents: theory and practice 143 accelerator control (Jennings et al., 1993), intelligent document retrieval (Mukhopadhyay et al., 1986), patient care (Huang et al., 1995), telecommunications network management (Weihmayer & Velthuijsen, 1994), spacecraft control (Schwuttke & Quan, 1993), computer integrated manufac turing (Parunak, 1995), concurrent engineering (Cutkosky et al., 1993), transportation manage ment (Fischer et al., 1993), job shop scheduling (Morley & Schelberg, 1993), and steel coil processing control (Mori et al., 1988). The classic reference to DAI is Bond and Gasser (1988), which includes both a comprehensive review article and a collection of significant papers from the field; a more recent review article is Chaib-draa et al. (1992). 5.2 Interface agents Macs defines interface agents as: "[C]omputer programs that employ artificial Intelligence techniques in order to provide assistance to a user dealing with a particular application .... The metaphor is that of a personal assistant who is collaborating with the user in the same work environment." (Macs, 1994h, p. 71) There are many interface agent prototype applications: for example, the New T system is a USENET news tilter (along the lines mentioned in the second scenario that introduced this article) (Maes, 1994a, pp. 38-39). A Ncw T agent is trained by giving it series of examples, illustrating articles that the user would and would not choose to read. The agent then begins to make suggestions to the user, and is given feedback on its suggestions. New T agents are not intended to remove human choice, but to represent an extension of the human's wishes: the aim is for the agent to be able to bring to the attention of the user articles of the type that the user has shown a consistent interest in. Similar ideas have been proposed by Mc Gregor, who imagines prescient agents-intelligent administrative assistants that predict our actions, and carry out routine or repetitive administrative procedures on our behalf (Mc Gregor, 1992). There is much related work being done by the computer supported cooperative work (CSCW) community. CSCW is informally defined by Haecker to be "computer assisted coordinated activity such as problem solving and communication carried out by a group of collaborating individuals" (Haecker, 1993, p. l). The primary emphasis of CSCW is on the development of (hardware and) software tools to support collaborative human work-the term gruupware has been coined to describe such tools. Various authors have proposed the use of agent technology in groupware. For example, in his participant systems proposal, Chang suggests systems in which humans collaborate with not only other humans, but also with artificial agents (Chang, 1987). We refer the interested reader to the collection of papers edited by Haecker (1993) and the article by Greif (1994) for more details on CSCW, 5.3 Information agents and cooperative information systems An information agent is an agent that has access to at least one, and potentially many information sources, and is able to collate and manipulate information obtained from these sources to answer queries posed by users and other information agents (the network of interoperating information sources are often referred to as intelligent and cooperative information systems (Papazoglou et al., 1992)). The information sources may be of many types, including, for example, traditional databases as well as other information agents. Finding a solution to a query might involve an agent accessing information sources over a network. A typical scenario is that of a user who has heard about somebody at Stanford who has proposed something called agent-oriented programming, The agent is asked to investigate, and, after a careful search of various FTP sites, returns with an appropriate technical report, as well as the name and contact details of the researcher involved. A number of studies have been made of information agents, including a theoretical study of how agents are able to incorporate information from different sources (Levy et al., 1994; Gruber, 1991 ), as well a prototype system called IRA (information retrieval agent) that is able to search for loosely

M. WOOLDRIDGE A:-1D NICHOLAS JENNINGS 142 agency discussed in this paper) is particularly important, as it potentially makes agent technology available to a user base that is industrially (rather than academically) oriented. While the development of various languages for agent-based applications is of undoubted importance, it is worth noting that all of the academically produced languages mentioned above are in some sense prototypes. Each was designed either to illustrate or examine some set of principles, and these languages were not, therefore, intended as production tools. Work is thus needed, both to make the languages more robust and usable, and to investigate the usefulness of the concepts that underpin them. As with architectures, work is needed to investigate the kinds of domain for which the different languages are appropriate. Finally, we turn to the relationship between an agent language and the corresponding theories that we discussed in section 2. As with architectures, it is possible to divide agent languages into various different categories. Thus AGENT0, PLACA, Concurrent Metate M, APRIL, and MAIL arc deliberative languages, as they arc all based on traditional symbolic AI techniques. ABLE, on the other hand, is a purely reactive language. With AGENT0 and PLACA, there is a clear (if informal) relationship between the programming language and the logical theory the language is intended to realise. In both cases, the programming language represents a subset of the corresponding logic, which can be interpreted directly. However, the relationship between logic and language is not formally defined. Like these two languages, Concurrent Metate M is intended to correspond to a logical theory. But the relationship hetween Concurrent Metate M and the corresponding logic is much more closely defined, as this language is intended to be a directly executable version of the logic. Agents in Concurrent Metate M, however, are not defined in terms of mentalistic constructs. For a discussion on the relationship between Concurrent Metate M and AGENT0-like languages, see Fisher (1995). 4.2 Further reading A recent collection of papers on concurrent object systems is Agha ct al. (1993). Various languages have been proposed that marry aspects of object-based systems with aspects of Shoham·s agent oriented proposal. Two examples are AGENTSPEAK and DAISY. AGENTSPEAK is loosely based on the PRS agent architecture, and incorporates aspects of concurrent-object technology (Weerasooriya et al., 1995). In contrast, DAISY is based on the concurrent-object language CUBL (Adorni & Poggi, 1993), and incorporates aspects of the agent-oriented proposal (Poggi, 1995). Other languages of interest include OZ (Henz et al., 1993) and IC PRO LOG Il (Chu, 1993). The latter, as its name suggests, is an extension of PROLOG, which includes multiple-threads, high-level communication primitives, and some object-oriented features. 5 Applications Although this article is not intended primarily as an applications review. it is nevertheless worth pausing to examine some of the current and potential applications of agent technology. 5.1 Cooperative problem solving and distributed Al As we observed in section 1, there has been a marked flowering of interest in agent technology since the mid-1980s. This interest is in part due to the upsurge of interest in Distributed Al Although DAI encompasses most ofthc issues we have discussed in this paper, it should be stressed that the dassical emphasis in DAI has been on macro phenomena (the social level), rather than the micro phenomena (the agent level) that we have been concerned with in this paper. DAI thus looks at such issues as how a group of agents can be made to cooperate in order to efficiently solve problems, and how the activities of such a group can be efficiently coordinated. DAI researchers have applied agent technology in a variety of areas. Example applications include power systems management (Wittig, 1992; Varga et al., 1994), air-traffic control (Steeb et al., 1988), particle

Intelligent agents: theory and practice 141 designed to provide the core features required to realise most agent architectures and systems. Thus APRIL provides facilities for multi-tasking (via processes, which are treated as first-class objects, and a Unix-like fork facility), communication (with powerful message-passing facilities supporting network-transparent agent-to-agent links); and pattern matching and symbolic process ing capabilities. The generality of APRIL comes at the expense of powerful abstractions-an APRIL system builder must implement an agent or system architecture from scratch using APRIL's primitives. In contrast, the MAIL language provides a rich collection of pre-defined abstractions, including plans and multi-agent plans. APRIL was originally envisaged as the implementation language for MAIL. The MAIL system has been used to implement several prototype multi-agent systems, including an urban traffic management scenario (Haugeneder and Steiner, 1994). 4.0.6 General Magic, lnc.-TELESCRIPT TELESCRIPT is a language-based environment for constructing agent societies that has been developed by General Magic, Inc.: it is perhaps the first commercial agent language. TELESCRIPT technology is the name given by General Magic to a family of concepts and techniques they have developed to underpin their products. There are two key concepts in TELESCRIPT technology: places and agents. Places are virtual locations that are occupied by agents. Agents are the providers and consumers of goods in the electronic marketplace applications that TELESCRIPTwas developed to support. Agents are software processes, and are mobile: they are able to move from one place to another, in which case their program and state are encoded and transmitted across a network to another place, where execution recommences. Agents are able to communicate with one-another: if they occupy different places, then they can connect across a network, in much the standard way; if they occupy the same location, then they can meet one another. Four components have been developed by General Magic to support TELESCRIPT tech nology. The first is the TELESCRIPT language. This language "is designed for carrying out complex communication tasks: navigation, transportation, authentication, access control, ·and so on" (White, 1994, p.17). The second component is the TELESCRIPT engine. An engine acts as an interpreter for the TELESCRIPT language, maintains places, schedules agents for execution, manages communication and agent transport, and finally, provides an interface with other applications. The third component is the TELESCRIPT protocol set. These protocols deal primarily with the encoding and decoding of agents, to support transport between places. The final component is a set of software tools to support the development of TELESCRIPT applications. 4.0.7 Connah and Wavish-ABLE A group at Philips research labs in the UK have developed an Agent Behaviour Language (ABLE), in which agents are programmed in terms of simple, rule-like licences (Connah & Wavish, 1990; Wavish, 1992). Licences may include some representation of time (though the language is not based on any kind of temporal logic): they loosely resemble behaviours in the subsumption architecture (see above). ABLE can be compiled down to a simple digital machine, realised in the "C" programming language. The idea is similar to situated automata, though there appears to be no equivalent theoretical foundation. The result of the compilation process is a very fast implementation, which has been used to control a Compact Disk-Interactive (CD-I) application. ABLE has recently been extended to a version called Real-Time ABLE (RTA) (Wavish & Graham, 1995). 4.1 Discussion The emergence of various language-based software tools for building agent applications is clearly an important development for the wider acceptance and use of agent technology. The release of TELESCRIPT, a commercial agent language (albeit one that does not embody the strong notion of

M. WOOLDRIDGE AND NICHOLAS JENNINGS 140 = CA~ open(door)8 Bl CA~ open (door)8. This formula is read: "if at time 5 agent a can ensure that the door is open at time 8, then at time 5 agent b believes that at time 5 agent a can ensure that the door is open at time 8". Corresponding to the logic is the AGENT0 programming language. In this language, an agent is specified in terms of a set of capabilities (things the agent can do), a set of initial beliefs and commitments, and a set of commitment rules. The key component, which determines how the agent acts, is the commitment rule set. Each commitment rule contains a message condition, a mental condition, and an action. To determine whether such a rule fires, the message condition is matched against the messages the agent has received; the mental condiiion is matched against the beliefs of the agent. If the rule fires, then the agent becomes committed to the action. Actions may be private, corresponding to an internally executed subroutine, or communicative, i.e., sending messages. Messages are constrained to be one of three types: "requests" or "unrequests" to perform or refrain from actions, and "inform" messages, which pass on information-Shoham indicates that he took his inspiration for these message types from speech act theory (Searle, 1969; Cohen & Perrault, 1979). Request and unrequest messages typically result in the agent's commitments being modified; inform messages result in a change to the agent's beliefs. 4.0.3 Tltomas-PLACA AGENT0 was only ever intended as a prototype, to illustrate the principles of AOP. A more refined implementation was developed by Thomas, for her 1993 doctoral thesis (Thomas, 1993). Her Planning Communicating Agents (PLACA) language was intended to address one severe drawback to AGENT0: the inability of agents to plan, and communicate requests for action via high-level goals. Agents in PLACA are programmed in much the same way as in AGENT0, in terms of mental change rules. The logical component of PLACA is similar to AGENTO's, but includes operators for planning to do actions and achieve goals. The semantics of the logic and its properties are examined in detail. However, PLACA is not at the "production" stage; it is an experimental language. 4.0.4 Fisher-Concurrent Metate M One drawback with both AGENT0 and PLACA is that the relationship between the logic and interpreted programming language is only loosely defined: in neither case can the programming language be said to truly execute the associated logic. The Concurrent Metate M language developed by Fisher can make a stronger claim in this respect (Fisher, 1994). A Concurrent Metate M system contains a number of concurrently executing agents. each of which is able to communicate with its peers via asynchronous broadcast message passing. Each agent is pro grammed by giving it a temporal logic specification of the behaviour that it is intended the agent should exhibit. An agent's specification is executed directly to generate its behaviour. Execution of the agent program corresponds to iteratively building a logical model for the temporal agent specification. It is possible to prove that the procedure used to execute an agent specification is correct, in that if it is possible to satisfy the specification, then the agent will do so (Barringer et al., 1989). The logical semantics of Concurrent Metate M are closely related to the semantics of temporal logic itself. This means that, amongst other things, the specification and verification of Concurrent Metate M systems is a realistic proposition (Fisher & Wooldridge, 1993). At the time of writing, only prototype implementations of the language are available; full implementations are expected soon. 4.0.5 The IMAGINE Project-APRIL and MAIL ,APRIL (Mc Cabe & Clark, 1995) and MAIL (Haugeneder ct al., 1994) are two languages for developing multi-agent applications that were developed as part of the ESPRIT project IMAGINE (Haugenedcr, 1994). The two languages arc intended to fulfil quite different roles. APRIL was

Intelligent agents: theory and practice 139 plans, which are essentially decision trees that can be used to efficiently determine an appropriate action in any situation (Schoppers, 1987). Another proposal for building "reactive planners" involves the use of reactive action packages (Firby, 1987). Other hybrid architectures are described in Hayes-Roth (1990), Downs and Reichgelt (1991), Aylett and Eustace (1994) and Bussmann and Demazeau (1994). 4 Agent languages As agent technology becomes more established, we might expect to see a variety of software tools become available for the design and construction of agent-based systems; the need for software support tools in this area was identified as long ago as the mid-1980s (Gasser et al., 1987). The emergence of a number of prototypical agent languages is one sign that agent technol0gy is becoming more widely used, and that many more agent-based applications are likely to be developed in the near future. By an agent language, we mean a system that allows one to program hardware or software computer systems in terms of some of the concepts developed hy agent theorists. At the very least, we expect such a language to include some structure corresponding to an agent. However, we might also expect to sec some other attributes of agency (beliefs, goals, or other mentalistic notions) used to program agents. Some of the languages we consider below embody this strong notion of agency; others do not. However, all have properties that make them interesting from the point of view of this review. 4.0. I Concurrent object languages Concurrent object languages are in many respects the ancestors of agent languages. The notion of a self-contained concurrently executing object, with some internal state that is not directly accessible to the outside world, responding to messages from other such objects, is very close to the concept of an agent as we have defined it. The earliest concurrent object framework was Hewitt's Actor model (Hewitt, 1977; Agha, 1986); another well-known example is the ABCL system (Yonezawa, 1990). For a discussion on the relationship between agents and concurrent object programming, see Gasser and Briot (1992). 4.0.2 Shoham-agent-oriented programming Yoav Shoham has proposed a "new programming paradigm, based on a societal view of computation" (Shoham, 1990, p. 4; 1993). The key idea that informs this agent-oriented program ming (AOP) paradigm is that of directly programming agents in terms of the mentalistic, intentional notions that agent theorists have developed to represent the properties of agents. The motivation behind such a proposal is that, as we observed in section 2, humans use the intentional stance as an abstraction mechanism for representing the properties of complex systems. ln the same way that we use the intentional stance to describe humans, it might be useful to use the intentional stance to program machines. Shoham proposes that a fully developed AOP system will have three components: • a logical system for defining the mental state of agents; • an interpreted programming language for programming agents; • an "agentification" process, for compiling agent programs into low-level executable systems. At the time of writing, Shoham has only published results on the first two components. (In Shoham (1990, p. 12), he wrote that "the third is still somewhat mysterious to me", though later in the paper he indicated that he was thinking along the lines of Rosenschcin and Kaelbling·s situated automata paradigm (Rosenschcin & and Kaelbling, 1986).) Shoham·s first attempt at an AOP language was the AGENT0 system. The logical component of this system is a quantified multi-modal logic, allowing direct reference to time. No semantics are given, but the logic appears to be based on Thomas et al. (1991). The logic contains three modalities: belief, commitment and ability. The following is an acceptable formula of the logic, illustrating its key properties:

M. WOOLDRIDGE AND NICHOLAS JENNINGS 138 relationships in Al, of which a particularly relevant example is Rao and Georgeff (1992a). This article discusses the relationship between the abstract BDI logics developed by Rao et al. for reasoning about agents, and an abstract "agent interpreter", based on the PRS. However, the relationship between the logic and the architecture is not formalised; the BDI logic is not used to give a formal semantics to the architecture, and in fact it is difficult to see how such a logic could he used for this purpose. A serious attempt to define the semantics of a (somewhat simple) agent architecture is presented in Wooldridge (1995), where a formal model of the system My World, in which agents are directly programmed in terms of beliefs and intentions, is used as the basis upon which to develop a logic for reasoning about My World systems. Although the logic contains modalities for representing beliefs and intentions, the semantics of these modalities are given in terms of the agent architecture itself, and the problems associated with possible worlds do not, therefore, arise; this work builds closely on Konolige's models of the beliefs of symbolic AI systems (Konoligc, 1986a). However, more work needs to be done using this technique to model more complex architectures, before the limitations and advantages of the approach are well-understood. Like purely deliberative architectures, some reactive systems are also underpinned by a relatively transparent theory. Perhaps the best example is the situated automata paradigm, where an agent is specified in terms of a logic of knowledge, and this specification is compiled down to a simple digital machine that can he realistically said to realise its corresponding specification. However, for other purely reactive architectures, based on more ad hoc principles, it is not clear that there is any transparent underlying theory. It could be argued that hybrid systems also tend to bead hoc, in that while their structures are well-motivated from a design point of view, it is not clear how one might reason about them, or what their underlying theory is. In particular, architectures that contain a number of independent activity producing subsystems, which compete with each other in real time to control the agent's activities, seem to defy attempts at formalisation. It is a matter of debate whether this needs he considered a serious disadvantage, but one argument is that unless we have a good theoretical model of a particular agent or agent architecture, then we shall never really understand why it works. This is likely to make it difficult to generalise and reproduce results in varying domains. 3.5 Further reading Most introductory textbooks on Al discuss the physical symbol system hypothesis; a good recent example of such a text is Ginsberg (1993). A detailed discussion of the way that this hypothesis has affected thinking in symbolic AI is provided in Shardlow (1990). There are many objections to the symbolic AI paradigm, in addition to those we have outlined above. Again, introductory textbooks provide the stock criticisms and replies. There is a wealth of material on planning and planning agents. See Georgeff (1987) for an overview of the state of the art in planning (as it was in 1987), Allen et al. (1990) for a thorough collection of papers on planning (many of the papers cited above are included), and Wilkins (1988) for a detailed description of SIPE, a sophisticated planning system used in a real-world application (the control of a brewery!) Another important collection of planning papers is Georgeff and Lansky (1986). The book by Dean and Wellman and the book by Allen et al. contain much useful related material (Dean and Wellman, 1991; Allen et al., 1991 ). There is now a regular international conference on planning; the proceedings of the first were published as Hendler (1992). The collection of papers edited by Maes (1990a) contains many interesting papers on alterna tives to the symbolic AI paradigm. Kaelbling (1986) presents a clear discussion of the issues associated with developing resource-bounded rational agents, and proposes an agent architecture somewhat similar to that developed by Brooks. A proposal by Nilsson for teleo reactive programs goal directed programs that nevertheless respond to their environment-is described in Nilsson (1992). The proposal draws heavily on the situated automata paradigm; other work based on this paradigm is described in Shoham (1990) and Kiss and Reichgelt (1992). Schoppcrs has proposed compiling plans in advance, using traditional planning techniques, in order to develop universal

Intelligent agents: theory and practice 137 model, various patterns of behaviour may be activated, dropped, or executed. As a result of Po B execution, the plan-based <;omponent and cooperation component may be asked to generate plans and joint plans respectively, in order to achieve the goals of the agent. This ultimately results in primitive actions and messages being generated by the world interface. 3.4 Discussion The deliberative, symbolic paradigm is, at the time of writing, the dominant approach in (D)AL This state of affairs is likely to continue, at least for the near future. There seem to be several reasons for this. Perhaps most importantly, many symbolic AI techniques (such as rule-based systems) carry with them an associated technology and methodology that is becoming familiar to mainstream computer scientists and software engineers. Despite the well-documented problems with symbolic AI systems, this makes symbolic AI agents (such as GRATE*, Jennings, 1993b) an attractive proposition when compared to reactive systems, which have as yet no associated methodology. The need for a development methodology seems to be one of the most pressing requirements for reactive systems. Anecdotal descriptions of current reactive systems implemen tations indicate that each such system must be individually hand-crafted through a potentially lengthy period of experimentation (Wavish and Graham, 1995). This kind of approach seems unlikely to be usable for large systems. Some researchers have suggested that techniques from the domain of genetic algorithms or machine learning might be used to get around these development problems, though this work is at a very early stage. There is a pressing need for research into the capabilities of reactive systems, and perhaps in particular to the types of application for which these types of system are best suited; some preliminary work has been done in this area, using a problem domain known as the Tile World (Pollack & Ringuette, 1990) With respect to reactive systems, Ferguson suggests that: "[T]he strength of purely non-deliberative architectures lies in their ability to exploit local patterns of activity in their current surroundings in order to generate more or less hardwired action responses .. for a given set of stimuli Successful operation using this method pre-supposes: i that the complete set of environmental stimuli required for unambiguously determining action sequences is always present and readily identifiable-in other words, that the agent's activity can be situationally determined; ii that the agent has no global task constraints ... which need to be reasoned about at run time; and iii that the agent's goal or desire system is capable of being represented implicitly in the agent's structure according to a fixed, pre-compiled ranking scheme." (Ferguson. 1992a, pp. 29-30} Hybrid architectures, such as the PRS, Touring Machines, Inte RRa P, and COSY, are currently a very active area of work, and arguably have some advantages over both purely deliberative and purely reactive architectures. However, an outstanding problem with such architectures is that of combining multiple interacting subsystems (deliberative and reactive) cleanly, in a well-motivated control framework. Humans seem to manage different levels of abstract behaviour with compari tive ease; it is not clear that current hybrid architectures can do so. Another area where as yet very little work has been done is the generation of goals and intentions. Most work in AT assumes that an agent has a single, well-defined goal that it must achieve. But if agents are ever to be really autonomous, and act pro-actively, then they must be able to generate their own goals when either the situation demands, or the opportunity arises. Some preliminary work in this area is Norman and Long (1995). Similarly, little work has yet been done into the management and scheduling of multiple, possibly conflicting goals; some preliminary work is reported in Dongha (1995). Finally, we turn to the relationship between agent theories and agent architectures. To what extent do the agent architectures reviewed above correspond to the theories discussed in section 2? What, if any, is the theory that underpins an architecture? With respect to purely deliberative architectures, there is a wealth of underlying theory. The close relationship between symbolic processing systems and mathematical logic means that the semantics of such architectures can often be represented as a logical system of some kind. There is a wealth of work establishing such

M. WOOLDRIDGE AND '.'JICHOLAS JENNINGS 136 layers, and in particular, to deal with conflicting action proposals from the different layers. The control framework does this by using control rules. 3.3.3 Burmeister et al.-COSY The COSY architecture is a hybrid BDI-architecture that includes elements of both the PRS and IRMA, and was developed specifically for a multi-agent testbed called DASEDIS (Burmeister & Sundermeyer; Haddadi, 1994). The architecture has five main components: (i) sensors; (ii) actuators; (iii) communications (iv) cognition; and (v) intention. The first three components are straightforward: the sensors receive non-communicative perceptual input, the actuators allow the agent to perform non-communicative actions, and the communications component allows the agent to send messages. Of the remaining two components, the intention component contains "long-term goals, attitudes, responsibilities and the like ... the control elements taking part in the reasoning and decision-making of the cognition component" (Haddadi, 1994, p. 15), and the cognition component is responsible for mediating between the intentions of the agent and its beliefs about the world, and choosing an appropriate action to perform. Within the cognition component is the knowledge base containing the agent's beliefs, and three procedural components: a script execution component, a protocol execution component, and a reasoning, deciding and reacting component. A script is very much like a script in Schan k's original sense: it is a stereotypical recipe or plan for achieving a goal. Protocols are stereotypical dialogues representing cooperation frameworks such as the contract net (Smith, 1980). The reasoning, deciding and reacting component is perhaps the key component in COSY. It is made up of a number of other subsystems, and is structured rather like the PRS and IRMA (see above). An agenda is maintained, that contains a number of active scripts. These scripts may be invoked in a goal-driven fashion (to satisfy one of the agent's intentions), or a data-driven fashion (in response to the agent's current situation). A filter component chooses between competing scripts for execution 3.3.4 Muller et al.~lnte RRa P Intc RRa P, like Ferguson's Touring Machines, is a layered architecture, with each successive layer representing a higher level of abstraction than the one below it (Millier & Pischel, 1994; Millier et al., 1995; Millier, 1994). In Inte RRa P, these layers are further subdivided into two vertical layers: one containing layers of knowledge bases, the other containing various control components, that interact with the knowledge bases at their level. At the lowest level is the world interface control component, and the corresponding world level knowledge base. The world interface component, as its name suggests, manages the interface between the agent and its environment, and thus deals with acting, communicating, and perception. Above the world interface component is the behaviour-based component. The purpose of this component is to implement and control the basic reactive capability of the agent. This component manipulates a set of patterns of behaviour (Po B). A Po B is a structure containing a pre-condition that defines when the Po B is to be activated, various conditions that define the circumstances under which the Po B is considered to have succeeded or failed, a post-condition (a la STRIPS (Fikes & Nilsson, 1971)), and an executable body, that defines what action should be performed if the Po B is executed. (The action may be a primitive, resulting in a call on the agent's world interface. or may involve calling on a higher-level layer to generate a plan.) Above the behaviour-based component in Tnte RRa P is the plan-based component. This component contains a planner that is able to generate single-agent plans in response to requests from the behaviour-based component. The knowledge-base at this layer contains a set of plans, including a plan library. The highest layer of Inte RRa P is the cooperation component. This component is able to generate joint plans, that satisfy the goals of a number of agents, by elaborating plans selected from a plan library. These plans arc generated in response to requests from the plan-based component. Control in lnte RRa P is both data- and goal-driven. Perceptual input is managed by the world interface, and typically results in a change to the world model. As a result of changes to the world

Intelligent agents: theory and practice 135 some kind of precedence over the deliberative one, so that it can provide a rapid response to important environmental events. This kind of structuring leads naturally to the idea of a layered architecture, of which Touring Machincs (Ferguson, 1992) and Inte RRa P (Muller & Pischel, 1994) are good examples. (These architectures are described below.) In such an architecture, an agent's control subsystems are arranged into a hierarchy, with higher layers dealing with information at increasing levels of abstraction. Thus, for example, the very lowest layer might map raw sensor data directly onto effector outputs, while the uppermost layer deals with long-term goals. A key problem in such architectures is what kind of control framework to embed the agent's subsystems in, to manage the interactions between the various layers. 3.3.1 Georgeff and Lansky-PRS One of the best-known agent architectures is the Procedural Reasoning System (PRS), developed by Georgeff and Lansky (1987). Like IRMA (see above), the PRS is a belief-desire-intention architecture, which includes a plan library, as well as explicit symbolic representations of beliefs, desires, and intentions. Beliefs are facts, either about the external world or the system's internal state. These facts are expressed in classical first-order logic. Desires are represented as system behaviours (rather than as static representations of goal states). A PRSplan library contains a set of partially-elaborated plans, called knowledge areas (KA), each of which is associated with an ini,ocation condition. This condition determines when the KA is to be actil'ated. KAs may be activated in a goal-driven or data-driven fashion; KAs may also be reactive, allowing the PRS to respond rapidly to changes in its environment. The set of currently active KAs in a system represent its intentions. These various data structures are manipulated by a system interpreter, which is responsible for updating beliefs, invoking KAs, and executing actions. The PRS has been evaluated in a simulation of maintenance procedures for the space shuttle, as well as other domains (Georgcff & lngrand, 1989). 3.3.2 Ferguson-Touring Machines For his 1992 Doctoral thesis, Ferguson developed the Touring Machines hybrid agent architecture (Ferguson, 1992a,b).5 The architecture consists of perception and action subsystems, which interface directly with the agent's environment, and three control layers, embedded in a control framework, which mediates between the layers. Each layer is an independent, activity-producing, concurrently executing process. The reactive layer generates potential courses of action in response to events that happen too quickly for other layers to deal with. It is implemented as a set of situation-action rules, in the style of Brooks' subsumption architecture (see above). The planning layer constructs plans and selects actions to execute in order to achieve the agent's goals. This layer consists of two components: a planner, and a focus of attention mechanism. The planner integrates plan generation and execution. and uses a library of partially elaborated plans, together with a topological world map, in order to construct plans that will accomplish the agent's main goal. The purpose of the focus of attention mechanism is to limit the amount of information that the planner must deal with, and so improve its efficiency. It does this by filtering out irrelevant information from the environment. The modelling layer contains symbolic representations of the cognitive state of other entities in the agent's environment. These models are manipulated in order to identify and resolve goal conflicts-situations where an agent can no longer achieve its goals, as a result of unexpected interference. The three layers are able to communicate with each other (via message passing), and are embedded in a control framework. The purpose of this framework is to mediate between the 5 1t is worth noting that Fergeson's thesis gives a good overview of the problems and issues associated with building rational, resource-bounded agents. Moreover. the description given of the Touring Machines architecture is itself extremely clear. We recommend it as a point of departure for further reading.

M. WOOLDRIDGE AND NICHOLAS JENNINGS 134 "[An agent} ... x iss aid to carry the information that p inw orld states, written s I= K(x,p), if for all world states in which x has the same value as it does ins, the proposition pis true." (Kae!bling & Rosenschcin, 1990, p. 36) An agent is specified in terms of two components: perception and action. Two programs are then used to synthesise agents: RULER is used to specify the perception component of an agent; GAPPS is used to specify the action component. RULER takes as its input three components: "[A j specification of the semantics of the [a gent's} inputs ("whenever bit 1 is on, it is raining"); a set of static facts ("whenever it is raining, the ground is wet"); and a specification of the state transitions of the world ("if the ground is wet, it stays wet until the sun comes out"). The programmer then specifies the desired semantics for the output ("if this bit is on, the ground is wet"), and the compiler ... [synthesises] a circuit whose output will have the correct semantics .... All that declarative '"knowledge" has been reduced to a very simple circuit." (Kaelb!ing, 1991, p. 86) The GAPPS program takes as its input a set of goal reduction rules (essentially rules that encode information about how goals can be achieved), and a top level goal, and generates a program that can be translated into a digital circuit to realise the goal. Once again, the generated circuit does not represent or manipulate symbolic expressions; all symbolic manipulation is done at compile time. The situated automata paradigm has attracted much interest, as it appears to combine the best elements of both reactive and symbolic, declarative systems. However, at the time of writing, the theoretical limitations of the approach are not well understood; there are similarities with the automatic synthesis of programs from temporal logic specifications, a complex area of much ongoing work in mainstream computer science (see the comments in Emerson, 1990). 3.2.4 Maes-Agent network architecture Pattie Maes has developed an agent architecture in which an agent is defined as a set of competence modules (Macs, 1989, 1990b, 1991 ). These modules loosely resemble the behaviours of Brooks' subsumption architecture (above). Each module is specified by the designer in terms of pre- and post-conditions (rather like STRIPS operators), and an activation level, which gives a real-valued indication of the relevance of the module in a particular situation. The higher the activation level of a module, the more likely it is that this module will influence the behaviour of the agent. Once specified, a set of competence modules is compiled into a spreading activation network, in which the modules are linked to one-another in ways defined by their pre-and post-conditions. For example, if module a has post-condition rp, and module b has pre-condition cp, then a and bare connected by a successor link. Other types of link include predecessor links and conflicter links. When an agent is executing, various modules may become more active in given situations, and may be executed. The result of execution may be a command to an effector unit, or perhaps the increase in activation !eve! of a successor module. There are obvious similarities between the agent network architecture and neural network architectures. Perhaps the key difference is that it is difficult to say what the meaning of a node in a neural net is; it only has a meaning in the context of the net itself. Since competence modules are defined in declarative terms, however, it is very much easier to say what their meaning is. 3.3 Hybrid architectures Many researchers have suggested that neither a completely deliberative nor completely reactive approach is suitable for building agents. They have argued the case for hybrid systems, which attempt to marry classical and alternative approaches. An obvious approach is to build an agent out of two (or more) subsystems: a deliberative one, containing a symbolic world model, which develops plans and makes decisions in the way proposed by mainstream symbolic AI; and a reactive one, which is capable of reacting to events that occur in the environment without engaging in complex reasoning. Often, the reactive component is given

Intelligent agents: theory and practice 133 1. Intelligent behaviour can be generated without explicit representations of the kind that symbolic AI proposes. 2. Intelligent behaviour can be generated without explicit abstract reasoning of the kind that symbolic AI proposes. 3. Intelligence is an emergent property of certain complex systems. Brooks identifies two key ideas that have informed his research: 1. Situatedness and embodiment: "Real" intelligence is situated in the world, not in disembodied systems such as theorem provers or expert systems. 2. Intelligence and emergence: "Intelligent" behaviour arises as a result of an agent's interaction with its environment. Also, intelligence is "in the eye of the beholder"; it is not an innate, isolated property. If Brooks was just a Dreyfus-style critic of AI, his ideas might not have gained much currency. However, to demonstrate his claims, he has built a number of robots, based on the suhsumption architecture. A subsumption architecture is a hierarchy of task-accomplishing behaviours. Each behaviour "competes" with others to exercise control over the robot. Lower layers represent more primitive kinds of behaviour (such as avoiding obstacles), and have precedence over layers further up the hierarchy. It should be stressed that the resulting systems are, in terms of the amount of computation they need to do, extremely simple, with no explicit reasoning of the kind found in symbolic AI systems. But despite this simplicity, Brooks has demonstrated the robots doing tasks that would be impressive if they were accomplished by symbolic AI systems. Similar work has been reported by Steels, who described simulations of "Mars explorer" systems, containing a large number of subsumption-architecture agents, that can achieve near-optimal performance in certain tasks (Steels, 1990). 3.2.2 Agre and Chapman-PENG! At about the same time as Brooks was describing his first results with the subsumption architecture, Chapman was completing his Master's thesis, in which he reported the theoretical difficulties with planning described above, and was coming to similar conclusions about the inadequacies of the symbolic AI model himself. Together with his co-worker Agre, he began to explore alternatives to the AI planning paradigm (Chapman & Agre, 1986). Agre observed that most everyday activity is ··routine" in the sense that it requires little-if any-new abstract reasoning. Most tasks, once learned, can be accomplished in a routine way, with little variation. Agre proposed that an efficient agent architecture could be based on the idea of ''running arguments". Crudely, the idea is that as most decisions are routine, they can be encoded into a low-level structure (such as a digital circuit), which only needs periodic updating, perhaps to handle new kinds of problems. His approach was illustrated with the celebrated PENGI system (Agre & Chapman, 1987). PENGI is a simulated computer game, with the central character controlled using a scheme such as that outlined above. 3.2.3 Rosenschein and Kaelhling-situated automata Another sophisticated approach is that of Rosenschein and Kaclbling (Rosenschein, 1985; Rosenschein & Kaelbling, 1986; Kaelbling & Rosenschcin, 1990; Kaelbling, 1991). In their situated automata paradigm, an agent is specified in declarative terms. This specification is then compiled down to a digital machine, which satisfies the declarative specification. This digital machine can operate in a provably time-bounded fashion; it does not do any symbol manipulation, and in fact no symbolic expressions arc represented in the machine at all. The logic used to specify an agent is essentially a modal logic of knowledge (see above). The technique depends upon the possibility of giving the worlds in possible worlds semantics a concrete interpretation in terms of the states of an automaton:

M. WOOLDRIDGE AND NICHOLAS JENNINGS 132 which monitors the environment in order to determine further options for the agent; a filtering process; and a deliberation process. The filtering process is responsible for determining the subset of the agent's potential courses of action that have the property of being consistent with the agent's current intentions. The choice between competing options is made by the deliberation process. The IRMA architecture has been evaluated in an experimental scenario known as the Tileworld (Pollack & Ringuette, 1990). 3.1.3 Vere and Bickmore-HOMER An interesting experiment in the design of intelligent agents was conducted by Vere and Bickmore (1990). They argued that the enabling technologies for intelligent agents are sufficiently developed to be able to construct a prototype autonomous agent, with linguistic ability, planning and acting capabilities, and so on. They developed such an agent, and christened it HOMER. This agent is a simulated robot submarine, which exists in a two-dimensional "Seaworld'', about which it has only partial knowledge. HOMER takes instructions from a user in a limited subset of English with about an 800 word vocubulary; instructions can contain moderately sophisticated temporal references. HOMER can plan how to achieve its instructions (which typically relate to collecting and moving items around the Seaworld), and can then execute its plans, modifying them as required during execution. The agent has a limited episodic memory, and using this, is able to answer questions about its past experiences. 3.2.4 Jennings-GRATE* GRATE* is a layered architecture in which the behaviour of an agent is guided by the mental attitudes of beliefs, desires, intentions and joint intentions (Jennings, 1993b). Agents are divided into two distinct parts: a domain level system and a cooperation and control layer. The former solves problems for the organisation; be it in the domain of industrial control, finance or transportation. The latter is a meta-level controller which operates on the domain level system with the aim of ensuring that the agent's domain level activities are coordinated with those of others within the community. The cooperation layer is composed of three generic modules: a control module which interfaces to the domain level system, a situation assessment module and a cooperation module. The assessment and cooperation modules provide an implementation of a model of joint responsibility (Jennings, 1992), which specifics how agents should act both locally and towards other agents whilst engaged in cooperative problem solving. The performance of a GRATE* community has been evaluated against agents which only have individual intentions, and agents which behave in a selfish manner, in the domain of electricity transportation management. A significant improvement was noted when the situation became complex and dynamic (Jennings, 1995). 3.2 Alternative approaches: reactive architectures As we observed above, there arc many unsolved (some would say insoluble) problems associated with symbolic Al. These problems have led some researchers to question the viability of the whole paradigm, and to the development of what are generally known as reactive architectures. For our purposes, we shall define a reactive architecture to be one that does not include any kind of central symbolic world model, and does not use complex symbolic reasoning. 3.2. l Brooks-behaviour languages Possibly the most vocal critic of the symbolic AI notion of agency has been Rodney Brooks, a researcher at MIT who apparently became frustrated by AI approaches to building control mechanisms for autonomous mobile robots. In a 1985 paper, he outlined an alternative architec ture for building agents, the so called subsumption architecture (Brooks, 1986). The review of alternative approaches begins with Brooks' work. In recent papers, Brooks (1990, 1991a,b) has propounded three key theses:

Intelligent agents: theory and practice 131 reasoning, have turned out to be extremely difficult (cf. the CYC project (Guba & Lenat, 1994)). The underlying problem seems to be the difficulty of theorem proving in even very simple logics, and the complexity of symbol manipulation algorithms in general: recall that first-order logic is not even decidable, and modal extensions to it (including representations of belief, desire, time, and so on) tend to be highly undecidable. Thus, the idea of building "agents as theorem provers"-what might be called an extreme logicist view of agency-although it is very attractive in theory, seems, for the time being at least, to be unworkable in practice. Perhaps more troubling for symbolic AI is that many symbol manuipulation algorithms of interest are intractable. lt seems hard to build useful symbol manipulation algorithms that will he guaranteed to terminate with useful results in an acceptable fixed time bound. And yet such algorithms seem essential if agents are to operate in any real-world. time-constrained domain. Good discussions of this point appear in Kaelbling (1986) and Russell and Wefald (1991). It is because of these problems that some researchers have looked to alternative techniques for building agents; such alternatives are discussed in section 3.2. First, however, we consider efforts made within the symbolic Al community to construct agents. 3.1.1 Planning agents Since the early 1970s, the AI planning community has been closely concerned with the design of artificial agents; in fact, it seems reasonable to claim that most innovations in agent design have come from this community. Planning is essentially automaticprngamming: the design of aeourse of action that, when executed, will result in the achievement of some desired goal. Within the symbolic AI community, it has long been assumed that some form of Al planning system will be a central component of any artificial agent. Perhaps the best-known early planning system was STRIPS (Fikes & Nilsson, 1971). This system takes a symbolic description of both the world and a desired goal state, and a set of action descriptions, which characterise the pre- and post-conditions associated with various actions. It then attempts to find a sequence of actions that will achieve the goal, by using a simple means-ends analysis. which essentially involves matching the post conditions of actions against the desired goal. The STRIPS planning algorithm was very simple, and proved to be ineffective on problems of even moderate complexity. Much effort was subsequently devoted to developing more effective techniques. Two major innovations were hierarchical and non-linear planning (Sacerdoti, 1974, 1975). However, in the mid 1980s, Chapman established some theoretical results which indicate that even such refined techniques will ultimately turn out to be unusable in any time-constrained system (Chapman, 1987). These results have had a profound influence on subsequent AI planning research; perhaps more than any other, they have caused some researchers to question the whole symbolic AI paradigm, and have thus led to the work on alternative approaches that we discuss in section 3.2. In spite of these difficulties, various attempts have been made to construct agents whose primary component is a planner. For example: the Integrated Planning, Execution and Monitoring (IPEM) system is based on a sophisticated non-linear planner (Ambros-Ingerson and Steel, 1988); Wood's AUTODRIVE system has planning agents operating in a highly dynamic environment (a traffic simulation) (Wood, 1993); Etzioni has built "soft bots" that can plan and act in a Unix environment (Etzioni et al., 1994); and finally, Cohen's PHOENIX system includes planner-based agents that operate in the domain of simulated forest fire management (Cohen et al., 1989). 3.1.2 Rratman, Israel and Pollack-IRMA In section 2, we saw that some researchers have considered frameworks for agent theory based on beliefs, desires, and intentions (Rao & Georgeff, 1991b). Some researchers have also developed agent architectures based on these attitudes. One example is the Intelligent Resource-hounded Machine Architecture (IRMA) (Bratman et a!., 1988). This architecture has four key symbolic data structures: a plan library, and explicit representations of beliefs, desires, and intentions. Addition ally, the architecture has: a reasoner, for reasoning about the world; a means-end analyser, for determining which plans might be used to achieve the agent's intentions; an opportunity analyser,

M. WOOLDRIDGE A:-!D NICHOi.AS JENNINGS 130 3 Agent architectures Until now, this article has been concerned with agent theory-the construction of formalisms for reasoning about agents, and the properties of agents expressed in such formalisms. Our aim in this section is to shift the emphasis from theory to practice. We consider the issues surrounding the construction of computer systems that satisfy the properties specified by agent theorists. This is the area of agent architectures. Maes defines an agent architecture as: "[A] particular methodology for building [agents]. It specifies how ... the agent can be decomposed into the construction of a set of component modules and how these modules should be made to interact. The total set of modules and their interactions has to provide an answer to the question of how the sensor data and the current internal state of the agent determine the actions ... and future internal state of the agent. An architecture encompasses techniques and algorithms that support this methodology'· (Maes, 1991, p.115) Kaelbling considers an agent architecture to be: "[A] specific collection of software (or hardware) modules, typically designated by boxes with arrows indicating the data and control flow among the modules. A more abstract view of an architecture is as a general methodology for designing particular modular decompositions for particular tasks." (Kaelbling, 1991, p.86) The classical approach to building agents is to view them as a particular type of knowledge-based system. This paradigm is known as symbolic Al: we begin our review of architectures with a look at this paradigm, and the assumptions that underpin it. 3.1 Classical approaches: deliberative architectures The foundation upon which the symbolic AI paradigm rests is the physical-symbol system hypothesis, formulated by Newell and Simon (1976). A physical symbol system is defined to be a physically realisable set of physical entities (symbols) that can be combined to form structures, and which is capable of running processes that operate on those symbols according to symbolically coded sets of instructions. The physical-symbol system hypothesis then says that such a system is capable of general intelligent action. ft is a short step from the notion of a physical symbol system to Mc Carthy's dream of a sentential processing automaton, or deliberative agent. (The term "deliberative agent" seems to have derived from Genesercth"s use of the the term "deliberate agent'' to mean a specific type of symbolic architecture (Genesereth and Nilsson, 1987, pp. 325-327).) We define a deliberative agent or agent architecture to be one that contains an explicitly represented, symbolic model of the world, and in which decisions (for example about what actions to perform) arc made via logical (or at least pseudo-logical) reasoning, based on pattern matching and symbolic manipulation. The idea of deliberative agents based on purely logical reasoning is highly seductive: to get an agent to realise some theory of agency one might naively suppose that it is enough to simply give it logical representation of this theory and "get it to do a bit of theorem proving" (Shardlow. 1990, section 3.2). If one aims to build an agent in this way, then there are at least two important problems to be solved: 1. The transduction problem: that of translating the real world into an accurate, adequate symbolic description, in time for that description to be useful. 2. The representation/reasoning problem: that of how to symbolically represent information about complex real-world entities and processes, and how to get agents to reason with this information in time for the results to be useful. The former problem has led to work on vision, speech understanding, learning. etc. The latter has led to work on knowledge representation, automated reasoning, automatic planning, etc. Despite the immense volume of work that these problems have generated, most researchers would accept that neither is anywhere near solved. Even seemingly trivial problems, such as commonsense

330 G. E. M. ANSCOMBE motive. Plato saying to a slave ' I should beat you if I were not angry ' would be a case. Or a man might have a policy of never making remarks about a certain person because he could not speak about that man unenviously or unadmiringly. We have now distinguished between a backward-looking motive and a mental cause, and found that here at any rate what the agent reports in answer to the question ' Why? ' is a reason-for-acting if, in treating it as a reason, he conceives it as something good or bad, and his own action as doing good or harm. If you could e.g. showthat either the action for which he has revenged himself, or that in which he has revenged himself, was quite harmless or beneficial, he ceases to offer a reason, except prefaced by ' I thought '. If it is a proposed revenge he either gives it up or changes his reasons. No such discovery would affect an assertion of mental causality. Whether in general good and harm play an essential part in the concept of intention is something it still remains to find out. So far good and harm have only been introduced as making a clear difference between a backward-looking motive and a mental cause. When the question ' Why? ' about a present action is answered by description of a future state of affairs, this is already distinguished from a mental cause just by being future. Here there does not so far seem to be any need to characterise intention as being essentially of good or of harm. Now, however, let us consider this case: Why did you do it? Because he told me to. Is this a cause or a reason i It appears to depend very much on what the action was or what the circumstances were. And we should often refuse to make any distinction at all between something's being a reason and its being a cause of the kind in question; for that was explained as what one is after if one asks the agent what led up to and issued in an action, but being given a reason and accepting it might be This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

INTENTION 329 of ... often comes to the same as saying he did so lest ... or in order that . . . should not happen. Leaving then, the topic of motive-in-general or ' inter pretative ' motive, let us return to backward-looking motives. Why is it that in revenge and gratitude, pity and remorse, the past event (or present situation) is a reason for acting, not just a mental cause? Now the most striking thing about these four is the way in which good and evil are involved in them. E.g. if I am grateful to someone, it is because he has done me some good, or at least I think he has, and I cannot show gratitude by something that I intend to harm him. In remorse, I hate some good things for myself; I could not express remorse by getting myself plenty of enjoyments, or for something that I did not find bad. If I do something out of revenge which is in fact advantageous rather than harmful to my enemy, my action, in its description of being advantageous to him, is involuntary. These facts are the clue to our present problem. If an action has to be thought of by the agent as doing good or harm of some sort, and the thing in the past as good or bad, in order for the thing in the past to be the reason for the action, then this reason shows not a mental cause but a motive. This will come out in the agent's elaborations on his answer to the question 'Why?' It might seem that this is not the most important point, but that the important point is that a proposed action can be questioned and the answer be a mention of something past. 'I am going to kill him.'-' Why?'-' He killed my father.' But do we yet know what a proposal to act is; other than a prediction which the predictor justifies, if he does justify it, by mentioning a reason for acting? and the meaning of the expression ' reason for acting ' is precisely what we are at present trying to elucidate. Might one not predict mental causes and their effects? Or even their effects after the causes have occurred? E.g. 'This is going to make me angry.' Here it may be worth while to remark that it is a mistake to think one cannot choose whether to act from a 2 L This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

INTENTION may easily be inclined to deny both that there is any such thing as mental causality, and that ' motive ' means anything but intention. But both of these inclinations are mistaken. We shall create confusion if we do not notice (a) that phenomena deserving the name of mental causality exist, for we can make the question 'Why?' into a request for the sort of answer that I considered under that head; ( b) that mental causality is not restricted to choices or voluntary or intentional actions but is of wider application; it is restricted to the wider field of things the agent knows about not as an observer, so that it includes some involuntary actions; (c) that motives are not mental causes; and (d) that there is application for ' motive ' other than the applications of' the intention with which a man acts '. Revenge and gratitude are motives; if I kill a man as an act of revenge I may say I do it in order to be revenged, or that revenge is my object; but revenge is not some further thing obtained by killing him, it is rather that killing him is revenge. Asked why I killed him, I reply ' Because he killed my brother.' We might compare this answer, which describes a concrete past event, to the answer describing a concrete future state of affairs which we sometimes get in statements of objectives. It is the same with gratitude, and remorse, and pity for something specific. These motives differ from, say, love or curiosity or despair in just this way: something that has happened (o r is at present happening) is given as the ground of an action or abstention that is good or bad for the person (it may be oneself, as with remorse) at whom it is aimed. And ifwe wanted to explain e.g. revenge, we should say it was harming someone because he had done one some harm; we should not need to add some description of the feelings prompting the action or of the thoughts that had gone with it. Whereas saying that someone does something out of, say, friendship cannot be explained in any such way. I will call revenge and gratitude and remorse and pity backward-looking motives, and contrast them with motive-in-general. Motive-in-general is a very difficult topic which I do This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

326 G. E, M, ANSCOMBE him from this awful suffering ', or ' to get rid of the swine '; but though these are forms of expression suggesting objectives, they are perhaps expressive of the spirit in which the man killed rather than descriptive of the end to which the killing was a means-a future state of affairs to be produced by the killing. And this shows us part of the distinction that there is between the popular senses of motive and intention. We should say: popularly, ' motive for an action ' has a rather wider and more diverse application than ' intention with which the action was done '. When a man says what his motive was, speaking popu larly, and in a sense in which' motive ' is not interchangeable with ' intention ', he is not giving a ' mental cause ' in the sense that I have given to that phrase. The fact that the mental causes were such-and-such may indeed help to make his claim intelligible. And further, though he may say that his motive was this or that one straight off and without lying-i.e. without saying what he knows or even half knows to be untrue-yet a consideration of various things, which may include the mental causes, might possibly lead both him and other people to judge that his declaration of his own motive was false. But it appears to me that the mental causes are seldom more than a very trivial item among the things that it would be reasonable to consider. As for the importance of considering the motives of an action, as opposed to considering the intention, I am very glad not to be writing either ethics or literary criticism, to which this question belongs. Motives may explain actions to us; but that is not to say that they' determine', in the sense of causing, actions. We do say: ' His love of truth caused him to . . . ' and similar things, and no doubt such expressions help us to think that a motive must be what produces or brings about a choice. But this means rather ' He did this in that he loved the truth '; it interprets his action. Someone who sees the confusions involved in radically distinguishing between motives and intentions and in defining motives, so distinct, as the determinants of choice, This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

INTENTION thought or feeling in you? i.e., what did you see or hear or feel, or what ideas or images cropped up in your mind, and led up to it? I have isolated this notion of a mental cause because there is such a thing as this question with this sort of answer, and because I want to distinguish it from the ordinary senses of ' motive ' and 'intention', rather than because it is in itself of very great importance; for I believe that it is of very little. But it is important to have a clear idea of it, partly because a very natural conception of ' motive' is that it is what moves (the very word suggests that)-glossed as ' what causes ' a man's actions etc. And ' what causes ' them is perhaps then thought of as an event that brings the effect about-though how-i.e. whether it should be thought of as a kind of pushing in another medium, or in some other way-is of course completely obscure. In philosophy a distinction has sometimes been drawn between ' motives' and ' intentions in acting' as referring to quite different things. A man's intention is what he aims at or chooses; his motive is what determines the aim or choice; and I suppose that ' determines ' must here be another word for ' causes '. Popularly, ' motive ' and ' intention ' are not treated as so distinct in meaning. E.g. we hear of 'the motive of gain '; some philosophers have wanted to say that such an expression must be elliptical; gain must be the intention, and desire ofgain the motive. Asked for a motive, a man might say' I wanted to . . . 'which would please such philosophers; or 'I did it in order to ... ' which would not; and yet the meaning of the two phrases is here identical. When a man's motives are called good, this may be in no way distinct from calling his intentions good-e.g. ' he only wanted to make peace among his relations '. Nevertheless there is even popularly a distinction between the meaning of ' motive ' and the meaning of 'intention'. E.g. if a man kills someone, he may be said to have done it out of love and pity, or to have done it out of hatred; these might indeed be ·cast in the forms' to release This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

332 G. E. M. ANSCOMBE cases are the right ones to consider in order to see the distinction between reason and cause. But it is worth noticing that what is so commonly said, that reason and cause are everywhere sharply distinct notions, is not true. This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

324 G, E, M. ANSCOMBE saying that it mentions something future-this is also a case of a mental cause. For couldn't it be recast in the form: ' Because I wanted . . . ' or ' Out of a desire that . . . '? If a feeling of desire for an apple affects me and I get up and go to a cupboard where I think there are some, I might answer the question what led to this action by mentioning the desire as having made me ... etc. But it is not in all cases that ' I did so and so in order to . . . ' can be backed up by ' I felt a desire that . . . ' I may e.g. simply hear a knock on the door and go downstairs to open it without experiencing any such desire. Or suppose I feel an upsurge of spite against someone and destroy a message he has received so that he shall miss. an appointment. If I describe this by saying' I wanted to make him miss that appointment ', this does not necessarily mean that I had the thought ' If I do this, he will . . . ' and that it affected me with a desire of bringing that about which led up to my action. This may have happened, but need not. It could be that all that happened was this: I read the message, had the thought ' That unspeakable man ! ' with feelings of hatred, tore the message up, and laughed. Then if the question ' Why did you do that? ' is put by someone who makes it clear that he wants me to mention the mental causes-i.e., what went on in my mind and issued in the action-I should perhaps give this account; but normally the reply would be no such thing. That particular enquiry is not very often inade. Nor do I wish to say that it always has an answer in cases where it can be made. One might shrug or say 'I don't know that there was any definite history of the kind you mean ', or ' It merely occurred to me . . . ' A ' mental cause ', of course, need not be a mental event, i.e., a thought or feeling or image; it might be a knock on the door. But if it is not a mental event, it must be something perceived by the person affected-e.g. tthe knock on the door must be heard-so if in this sense anyone wishes to say it is always a mental event, I have no objection. A mental cause is what someone would describe if he were asked the specific question: what produced this action or This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

322 G, E, M. ANSCOMBE we cannot say 'Ah, but not a reason for acting;' we should be going round in circles. We need to find the difference between the two kinds of' reason' without talking about ' acting '; and if we do, perhaps we shall discover what is meant by ' acting' when it is said with this special emphasis. It will hardly be enlightening to say ' in the case of the sudden start the "reason'' is a cause'; the topic of causality is in a state of too great confusion; all we know is that this is one of the places where we do use the word ' cause '. But we also know that this is rather a strange case of causality; the subject is able to give a cause of a thought or feeling or bodily movement in the same kind of way as he is able to state the place of his pain or the position of his limbs. Such statements are not based on observation. Nor can we say: 'Well, the "reason" for a movement is a cause, and not a reason in the sense of "reason for acting ", when the movement is involuntary; it is a reason as opposed to a cause, when the movement is voluntary and intentional.' This is partly because in any case the object of the whole enquiry is really to delineate such concepts as the voluntary and the intentional, and partly because one can also give a ' reason ' which is only a ' cause ' for what is voluntary and intentional. E.g. ' Why are you walking up and down like that? ' - ' It's that military band; it excites me.' Or 'What made you sign the document at last?'-' The thought:" It is my duty" kept hammering away in my mind until I said to myself" I can do no other", and so signed.' Now we can see that the cases where this difficulty arises are just those where the cause itself, qua cause, (or perhaps one should rather say the causation itself) is in the class of things known without observation. I will call the type of cause in question a ' mental cause '. Mental causes are possible, not only for actions (' The martial music excites me, that is why I walk up and down ') but This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

Meeting of the Aristotelian Society at 21, Bedford Square, London, W.C. l, on 3rd June, 1957, at 7.30 p.m. XIV.-INTENTION By G. E. M. ANSCOMBE What distinguishes actions which are intentional from those which are not? The answer that suggests itself is that they are the actions to which a certain sense of the question 'Why?' is given application; the sense is defined as that in which the answer, if positive, gives a reason for acting. But this hardly gets us any further, because the questions ' What is the relevant sense of the question " Why? " ' and ' What is meant by " re;;i.son for acting " ? ' are one and the same. To see the difficulties here, consider the question ' Why did you knock the cup off the table?' answered by' I thought I saw a face at the window and it made me jump.' Now we cannot say that since the answer mentions something previous to the action, this will be a cause as opposed to a reason; for if you ask ' Why did you kill him? ' the answer ' he killed my father ' is surely a reason rather than a cause, but what it mentions is previous to the action. It is true that we don't ordinarily think of a case like giving a sudden start when we speak of a reason for acting. ' Giving a sudden start ', someone might say, ' is not acting in the sense suggested by the expression "reason for acting".' Hence, though indeed we readily say e.g. ' What was the reason for your starting so violently? ' this is totally unlike ' What is your reason for excluding so-and-so from your will? ' or ' What is your reason for sending for a taxi? ' But what is the difference ? Why is giving a start or gasp not an ' action ', while sending for a taxi or crossing the road is one? The answer cannot be ' Because an answer to the question " why? " may give a reason in the latter cases ', for the answer may ' give a reason ' in the former cases too; and 2K This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

INTENTION 331 such a thing. And how would one distinguish between cause and reason in such a case as having hung one's hat on a peg because one's host said ' Hang up your hat on that peg '? Nor, I think, would it be correct to say that this is a reason and not a mental cause because of the understanding of the words that went into obeying the suggestion. Here one would be attempting a contrast between this case and, say, turning round at hearing someone say Boo ! But this case would not in fact be decisively on one side or the other; forced to say whether the noise was a reason or a cause, one would probably decide by how sudden one's reaction was. Further, there is no question of understanding a sentence in the following case: ' Why did you waggle your two fore-fingers by your temples?'-' Becasue he was doing it; ' but this is not particularly different from hanging one's hat up because one's host said ' Hang your hat up.' Roughly speaking, if one were forced to go on with the distinction, the more the action is described as a mere response, the more inclined one would be to the word ' cause '; while the more it is described as a response to something as having a significance that is dwelt on by the agent, or as a response surrounded with thoughts and questions, the more inclined one would be to use the word ' reason '. But in very many cases the distinction would have no point. This, however, does not mean that it never has a point. The cases on which we first grounded the distinction might be called ' full-blown ': that is to say, the case of e.g. revenge on the one hand, and of the thing that made me jump and knock a cup off a table on the other. Roughly speaking, it establishes something as a reason to object to it, not as when one says ' Noises should not make you jump like that: hadn't you better see a doctor? ' but in such a way as to link it up with motives and intentions. ' You did it because he told you to? But why do what he says? ' Answers like ' he has done a lot for me '; ' he is my father '; ' it would have been the worse for me if I hadn't ' give the original answer a place among reasons. Thus the full-blown This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

INTENTION 323 also for feelings and even thoughts. In considering actions, it is important to distinguish between mental causes and motives; in considering feelings, such as fear or anger, it is important to distinguish between mental causes and objects of feeling. To see this, consider the following cases: A child saw a bit of red stuff on a turn in a stairway and asked what it was. He thought his nurse told him it was a bit of Satan and felt dreadful fear of it. (No doubt she said it was a bit of satin.) What he was frightened of was the bit of stuff; the cause of his fright was his nurse's remark. The object of fear may be the cause of fear, but, as Wittgenstein1 remarks, is not as such the cause of fear. (A hideous face appearing at the window would of course be both cause and object, and hence the two are easily confused.) Or again, you may be angry at someone's action, when what makes you angry is some reminder of it, or someone's telling you of it. This sort of cause of a feeling or reaction may be reported by the person himself, as well as recognised by someone else, even when it is not the same as the object. Note that this sort of causality or sense of ' causality ' is so far from accommodating itself to Hume's explanations that people who believe that Hume pretty well dealt with the topic of causality would entirely leave it out of their calculations; if their attention were drawn to it they might insist that the word 'cause' was inappropriate or was quite equivocal. Or conceivably they might try to give a Humeian account of the matter as far as concerned the outside observer's recognition of the cause; but hardly for the patient's. Now one might think that when the question ' Why? ' is answered by giving the intention with which a person acts-a case of which I will here simply characterise by 1 Philosophical Investigations, SS4776. This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

328 G. E. M. ANSCOMBE not want to discuss at any length. Consider the statement that one motive for my signing a petition was admiration for its promoter, X. Asked 'Why did you sign it?' I might well say' Well, for one thing, X, who is promoting it, did ... ' and describe what he did in an admiring way. I might add ' Of course, I know that is not a ground for signing it, but I am sure it was one of the things that most influenced me '-which need not mean:' I thought explicitly of this before signing.' I say ' Consider this ' really with a view to saying ' let us not consider it here.' It is too complicated. The account of motive popularised by Professor Ryle does not appear satisfactory. He recommends construing ' he boasted from vanity ' as saying ' he boasted ... and his doing so satisfies the law-like proposition that whenever he finds a chance of securing the admiration and envy of others, he does whatever he thinks will produce this admiration and envy.'2 This passage is rather curious and roundabout in its way of putting what it seems to say, but I can't understand it unless it implies that a man could not be saidto have boasted from vanity unless he always behaved vainly, or at least very often did so. But this does not seem to be true. To give a motive (o f the sort I have labelled ' motive-in general ', as opposed to backward-looking motives and intentions) is to say something like ' See the action in this light.' To explain one's own actions by an account indica ting a motive is to put them in a certain light. This sort of explanation is often elicited by the question ' Why? ' The question whether the light in which one so puts one's action is a true light is a notoriously difficult one. The motives admiration, curiosity, spite, friendship, fear, love of truth, despair and a host of others are either of this extremely complicated kind, or are forward-looking or mixed. I call a motive forward-looking if it is an intention. For example, to say that someone did something for fear 2 The Concept of Mind, p. 89. This content downloaded from 132.174.234.36 on Fri, 05 Sep 2025 17:48:40 UTC All use subject to https://about.jstor.org/terms

AGENT AGE. Signifies thoee periods In the lives Incapacity for reproduction, eating In ei of persons of both sexes which enable them ther ses:, and whether arising from struc to do certain acts which, before they bad tural or other e&Utle8. arrived at those periods, they were prohibit· ed from doing. . AGEll'l'llm A. Sax. The true master The length of time during which a per or owner of a thing. Spelman. son has lived or a thing has existed. In the old books, "age" ls commouly used AGE.NHI.N.a.. In Saxon law. A guest to signify "foll age;" that le, the age of at an Inn, who, ha ,1ng stayed there for twenty-one years. Litt. I 259. three nights, was then accounted one of the -Le JPll age. The age at which the person family. Oowell. acquires full capacity to make his own con tracts and deeds and transact business general• ly (age of majority) or to enter into some par AGEXB. Lat. An agent, a conductor. ticular contract or relation, u, the "legal age or manager of atralrs. Dlstlngulsbed from of consent" to marriage. See Capwell v. Cap factor, a workman. A plalutltr. Fleta, lib. well, 21 R. I. 101, 41 Atl. 1005: Montoya de 4, c. 15, I 8. Antonio v, Miller, 7 N. M. 289, 84 Pac. 40, 21 L.R. A.699. ' AGElf T. One who represents and acts AGE, Awe, Aln. L. Fr. Water. Kel• for another under the contract or relation ham. of agency, q. t1. AGE PBAYEB. A suggestion of DOD• 1p m eci - o . l u . .o A a t g l e o n a er . a l A a ge i n e t n 11 t a is r e o e n i e th e e r m g p e lo tt y 6 e " d 0 l o in r age, made by an Infant party to a real ac his capacity as a profesalonal man or master Uon, with a prayer that the proceedings of an art or trade, or one to whom the principal may be deferred until his full age. It ls confides hie whole bnainesa or all transactions or functions of a desi,tnated class ; or be ls a now abolished. St. 11 Geo. IY.; 1 Wm. IV. person who la authorized b:, hia principal to c. 37, I 10; 1 Lil. Reg. M; S Bl Comm. 800. execute all deeds, sign all contracts, or pur chase all good1, required in a J:)llrticular trade, AGENCY A relation, created either by business, or employment. See Story, Ag. I 17; Butler v. Maples, 9 Wall. 706, 19 L. Ed. 822; express or Implied contract or by law, where Jaquea v. Todd, 3 Wend. (N. Y.) 00; Spri~g by one party (called the principal or con• field Engine Co. v. Kenned]', 7 Ind. App. 502, stltuent) delegates the transaction of some 84 N. E. 866;_ .Cruzan v. Smith. 41 Ind. 297; lawful business or the authority to do cer Godshaw v. .-:Jtruck, 109 Ky. 285, 58 S. W. 781, 61 L. R. A. 668. A BJ)e Cial agent le one tain acts for him or In relation to his rights employed to conduct a -particular transaction or or property, with more or less discretionary piece of business for hi• principal or authoris• power, to another person (called the agent, ed to perform a 11peclfied act. Bryant v. Moore, attorney, proxy, or delegate) who under 26 Me. 87, 46 Am. Dec. 00: Gibson v. Snow Hardware Co., 94 Ala. 346, 10 South. 304: takes to manage the affair and render him Coolei v. Perrine, 41 N. J. Law, 325, 32 Am. an account thereof. • State v. Hubbard, 58 Rep, -10. Kan. 797, 61 Pac. 290, 39 L. R. A. 860; Agents employed for the aale of goods or mel' chandlse are called "mercantile agents," and Sternaman v. Iusurance Co., 170 N. Y. 13, are of two principal claeses.-brokere and fac 62 N. E. 763, 57 L. R. A. 318, 88 Am. St. tors, (q. 11.;) a factor is sometimes called a Rep. 625; Wynegar v. State, 157 Ind. 577, "commission agent," or "commission merchant." 62 N. E. 38. Ruu. Mere. ~~ 1. 8:,aoa:,ma. The term "agent" la to be The contract of agency may be defined to be distinguished from it• 11ynonyms "servant," acontract by which one of the contracting par- "representative," and "tru11tee." A servant acts ties confides the management of some affair, to In behalf of hi, master and under the latter's be transacted on his account, to the other par direction and authority, but is re,tarded as a ty, who undertakes to do the business and renmere instrument, and not as the substitute or der an account of it. 1 Liverm. Prln. & Ag. 2. proxy of the master. Tumer v. Cro BB, 83 Tex. A contract by which one person, with greater 218, 18 S. W. 578, 16 L. R. A. 262; People or lees discretionary power, undertakes to rep resent an.other In certain business relations. v. Treadwell, 69 Cal. 226, 10 Pac. 502. A representative (such as an executor or an as• Whart. Ag. 1. algnee in bankruptcy) owe11 his power and au A relation between two or more persons, by thority to the law. whkb puts him In the place which one party, usually called the agent or of the person represented, although the latter attomey. is authorized to do certain acts for, or may have designated or chosen the represent•· in relation to the rights or property of the other, who 111 denominated the principal, con- tlve. A trustee acta in the interest and for the stituent, or employer. Bouvier. benefit of one person, but by an authority de rived from another person. -Agency deed of. A revocable and volun• tary trust for payment of debts. Wharton.- Agency of necessity A term sometimes ap- Ia lateraatloaal law; A diplomatic plied to the kind of implied agency which en agent is a person employed by a sovereign ables a wife to procure what is reasonably to manage his private atralrs, or those of his necessary for her maintenance and support on her hu11band's credit and at his expense, when subjects In his name, at the court of a for he fails to make proper provision for her neces- eign government. Wollf, Inst. Nat. I 1287, sities. Bostwick v. Brower, 22 Misc. Rep. 700, 49 N. Y. Supp. 1046. Ia the praotloe of the hoaae of lol Nla -• prh7 oo-ou. In a JJpeala, solicitors AGEXESIA. In medical Jurisprudence. and other peraona admitted to practlee In Impotentta pnerandl ; sexual Impotence; thoee courta In a Blmllar capacl to tl'.tt ot oogle Digitized by

AGENT 51 AG ILLAR IUS IOllcltors In ordinary court.a, are technically AGGRAVATIOl'f. Any clrcnmstance at called "agents." Macph. Prlv. Coun. 00. tending the commission of a crime or tort -Acent aacl patleat. A phrase Indicating which increases Its guilt or enormity or the state of a pel'IIOn who Is required to do a adds to Its injurious consequences, but which thin A'. and is at the same time the pel'IIOn to Is above and beyond the essential constitu whom it la done.-Looal ageat. One ap ents of the crime or tort itself. pointed to act as the rep?Hentative of a cor poration and transact its bu1ines1 cenerall1 Matter of &g Jtravatlon, -correctly understood, (or b UBlnea of a particular character) at a giv does not consist in acts of the &ame kind and en place or within a defined district. See Frick description aa th011e constituting the gist of the Co. v. Wright, 2S Tex. Civ. App. 340, M S. action, but In something done b1 the defendant, W. 608; Moore v. Freeman'• Nat. Bank, 92 on the occasion of committing the tresp&.88, N. C. 594-; Westem, etc., Organ Co. v. Andel' which is. to eome extent, of a different legal aon, 97 Tex. 432, 79 S. W. 517.-llaaadas character from the principal act complained of. aaeat. A pel'IIOn who is Invested with general Hathaway v. Rice, 19 Vt. 107. power, involving the exercise of judgment and di1eredon, as distinguished from an. ordinary Ia pleadbag. The Introduction of mat qent or emplo:,6, who acts in an inferior ca ter into the declaration which tends to In pacit.r.. and under the direction and control of crease the amount of damages, but does uot 1t1perior authority, both in regard to the extent of the work and the manner of executing the affect the right of action Itself. Steph. Pl. ame. Reddington v. Mariposa Land & Min. 257; 12 Mod. 597. Co.. 19 Hun (N. Y.) 406; Taylor v. Granite State Prov. Au'n, 136 N. Y. 343, 32 N. E. 992{ AGGREGATE. Composed of several ; 32 Am. St. Rep. 749; U. S. v. American Bel conststlng of mnny penons united together. Te L Co. (C. C.) 29 fed. 33: Up~r Mieaieaippl Tranap. Co. v. Whittaker, 16 Wis. 220: Fos 1 BJ. Comm. 469. ! t J e S r N v. . C W h . a r 9 l , e s 2 3 B L et . c h R e . r A L . u 4 m 0 b 0 e , r 4 C 9 o . a 1 m 5 . S S . t D . . R 5 ep 7 . , -Assresate eo:rpo-tio-♦ See Coa PORA· TION. 859.-Prl Tate ageat. An. agent acting for an p in i d a i h v e id d u a fr l o m in a h p i1 u bl _ i p c ri a v g a e te n t, a w lf h a o lr s r ; e pr a e s se n d t i s l t t i h n e AGGREGATIO .-E.NTIU-♦ The meet government in aome adminlatrative capacit:,. Ing of minds. The moment when a contract Pa•:U. ..-1:. An acent of the public, the 18. complete. A suppoeed derivation of the atate, or the government; a pe IIIOn appointed word "agreement." to act for the nubile in some matter pertaining to the administration of rovemment or the put,;. lie balrine BS. See Sto_ry, Ag. I 300; Whiteside AGGRESBOR. The party who first of v. United States. 98 U. S. 2G4. 23 L. Ed. 882. fers violence or offense. He who ·begins a -a.al-estate a,ieat. Any pel'IIOn whose quarrel or dispute, either 'by threatening or lmsineu it is to sell, or offer for aale, real ea- striking another. . tate for others, or to ren.t hout!letl, etoree, or other buildings, or real estate, or to collect ftllt for othera. Act .Jul1 13. 1866, c. 49; 14 AGGRIEVED. Having suffered 1088 or St. at Larire, 118. Carstens v. Mc Rea-q, 1 injury ; . damnlfied : Injured. Wub. St. 359, 25 Pac. 4TI. AGGRIEVED PARTY. Under statutes Aseatff ot oo-•tleatff part pe11- granting the right of appeal to the party ttleoteat1a. Acting and consenting parties aggrieved by an order or Judgment, the par are liable to the llll JDe punishment. CS Coke, ty aggrieved ls one whose pecuniary inter 8). est ls directly affected by the adjudleatlon; one whose right of property may be estab A.GEIL Lat. la tile olvil law. A. lished or divested thereby. Ruff v. Mont Geld; land generally. A portion of land in gomery, 83 MIBB. 185, 36 South. 67; Mc Far claeed by definite boundaries. Munlcipallty land v. Pierce, 161 Ind. M6, 45 N. E. 706; No. 2 v. Orleans Ootton Preee, 18 La. 167, 36 Lamar v. Lamar, 118 Ga. 684, 45 8. E. 498; Am. Dec.~ Smith v. Bradstreet, 16 Pick. (Mass.) 264; Bryant v. Allen, 6 N. H. 116; Wiggin v. Ia o W Eas Hu law. An acre. Spelman. Swett, 6 Mete. (Maes.) 194, 39 Am. Dec. 716 ; Tillinghast v. Brown Unlvenlty, 24 R. I. 179, AGGEIL Lat. In the civil law. A dam, 62 Atl. 891 ; Lowery v. Lowery, 64 N. C. bank or mound. Cod. 9, 38; Townsh. Pl. 110; Raleigh v. Rogers, 25 N . .J. Eq. 006. Or 48. one against whom error bas been commltte1l. Kinealy v. Macklin, 67 Mo. 93. AGGRAVATED AIJBA'IJLT. An as ault with circumstances of aggravation, or A.GILD. In Saxon law. Free from pen of a heinous character, or with Intent to alty, not subject to the payment of g Ud, or eommft another crime. In re Burns (C. C.) tcere Qild; • that is, the customary fine or pe 113 Fed. 99'l; Norton v. State, 14 Tex. 303. cuniary compensation for an offense. Spel See A.sli ULT. man; O>well Defined In Penn91lvanla u follows: "If any ))el'IIOn shall nnlawfally and malicloualy Inflict AGILER. In Saxon law. An ob8erver upon another penon, either with or without or Informer. &DJ weapon or instrument. an1 rrievo U11 bodil1 m o lla t e h rm a e n r ., o r I o J , 4 ' r ' l r B e u O t n D c l . , a w h B e fu r l i s l s h y h a t l l c l . u J. t b . e P s u t g a r u d b l . , l t .r D o . r l. o g w . f o p a u . n 4 d m 3 i 4 s a d , n e y I law A . f f D A J , h 4 ay J w UV ar ■ d . , b L e . r d L w il a t r . d, I n o r o l k d e e E p n e g r l ls o b f 1t7. the berd of cattle 1n a common Gfield. Cow;ell. oog e Digitized by