From: AAAI Technical Report WS-02-06. Compilation copyright © 2002, AAAI (www.aaai.org). All rights reserved. Decisions and Gamesfor BD Agents Mehdi Dastani Institute of Informationand Computing Sciences Utrecht University The Netherlands email: mehdi@cs.uu.nl Abstract Strategic gamesmodelthe interaction betweensimultaneous decisionsof a groupof agents.Thestarting pointof strategic games is a set of players(agents)havingcertainstrategies (decisions)andpreferencesonthe game’soutcomes.In this paperwedo not assumethe set of decisionsand preferences of agentsto be given,butderivethemfromtheir mentalattitudes. In particular,weintroducea rule-basedarchitecture for agentswithbeliefs anddesiresandexplainhowtheir decisions andpreferencescan be derived. Wespecify groups of suchagents,definea mapping fromtheir specificationto the specificationof the gametheyplay,anduse somefamiliar notionsfromgametheory,suchas ParetoefficiencyandNash equilibrium, to characterizethe interactionbetween their decisions. Wealso discussa reversemapping fromthe specificationof games that a groupof agentplayto the specifications of thoseagents.Thismapping can be usedto specifygroups of agentsthat canplaya ceratingame. Introduction One of the main problemsin agent theory is the distinction betweenformal theories and tools developedfor individual autonomous agents, and theories and tools developed for multi agent systems.In the social sciences, this distinction is called the micro-macrodichotomy.Theprototypical exampleis the distinction betweenclassical decision theory based on the expected utility paradigm(usually identified with the work of Neumannand Morgenstem(von Neumann &Morgenstern1944) and Savage(Savage 1954)) and classical gametheory (such as the workof Nashand morerecently the workof Axelrod). Whereasclassical decision theory is a kind of optimization problem(maximizingthe agent’s expectedutility), classical gametheory is a kind of equilibria analysis. There are several approachesin artificial intelligence, cognitive science, and practical reasoning(within philosophy) to describe the decision makingof individual agents and to bring the micro and macrodescriptions together. In these approaches, the decision makingof individual autonomous agents is describedin terms of other conceptsthan maximizing utility. For example,since the early 40sthere is a distinction betweenclassical decisiontheory and artificial Copyright©2002,American Associationfor Artificial Intelligence(www.aaai.org). All rights reserved. 37 Leendert van der Torte Department of Artificial Intelligence Vrije Universiteit Amsterdam The Netherlands email: torre@cs.vu.nl intelligence basedon utility aspiration levels and goal based planning (as pioneered by Simon (Simon 1981)). Moreover, in cognitive science and philosophythe decision making of individual agents is described in terms of concepts fromfolk psychologylike beliefs, desires and intentions. In these studies, the decision makingof individual agents is characterized in terms of a rational balance betweenthese concepts, and the decision makingof a group of agents is described in terms of conceptsgeneralized from those used for individual agents, such as joint goals, joint intentions, joint commitments,etc. Moreover,also new concepts are introducedat this social level, suchas norms(a central concept in mostsocial theories). It is still an open problem howthe micro-macro dichotomyof classical decision and gametheory is related to the micro-macrodichotomyof these alternative theoties. Has or can the micro and macro level be brought together by replacing classical theories by alternative theodes? Before this question can be answered, the relation betweenthe classical and alternative theories has to be clarified. Doyle and Thomason(Doyle & Thomasonsummer 1999) argue that classical decision theory should be reunited with alternative decision theories in so-called qualitative decisiontheory, whichstudies qualitative versionsof classical decision theory, hybrid combinationsof quantitative and qualitative approachesto decision making,and decision makingin the contextof artificial intelligence applications such as planning, learning and collaboration. Qualitative decision theories have been developedbased on beliefs (probabilities) anddesires (utilities) usingformaltools such as modallogic (Boutilier 1994)and on utility functions and knowledge(Lang 1996). Morerecently these beliefsdesires modelshave been extended with intentions or BDI models (Cohen & Levesque 1990; Rao & Georgeff 1991a; 1991b). In this paper, we introduce a rule based qualitative decision and gametheory for agents with beliefs and desires. Likeclassical decisiontheorybut in contrast to several proposals in the BDI approach (Cohen &Levesque 1990; Rao&Georgeff1991a), the theory does not incorporate decision processes, temporal reasoning, and scheduling. We also ignore probabilistic decisions and related gameswith mixedstrategies. In particular, we explain howdecisions and preferences of individual agents can be derived from their beliefs and desires. Wespecify groupsof agents and discuss the interaction betweentheir decisions and preferences. This enables us to define a mappingfrom agent specification to the specificationof the gamethat they play. This mappingconsiders agent decisions as agent strategies and the consequencesof decisions as the outcomesof the game. Finally, we discuss the reverse mappingfrom the specification of a gameto the specification of the agents that play the game.This mappingprovides the beliefs and desires of agents that can play a certain game. A qualitative decision and gametheory has the input T, is the extensioncalculated basedon _Rand T. Thisis formalizedin the followingdefinition. Definition 1 (Extension) Let R C_ LAWx LAWbe a set of rules and T c LAWbe a set of sentences. The consequents of the T-applicablerules are: R(T) = {y ] x ~ y ¯ R,x ¯ andthe R extension ofT is the set of the consequentsof the iteratively T-applicablerules: ER(T) ¢’~c_x,R(c,~aw(X))c_x X The qualitative decision and gametheory introducedin this section is developedfor agents that have conditional beliefs and desires, i.e. the beliefs and desires of agents are sets of rules. The architecture and the behavior of this type of agents are studied in (Broersen et al. 2001b; Dastani &van der Torre 2002). Here, we analyze this type of agents froma decision and gametheoretic point of view by studying possible decisions of individual agents and the interaction betweenthese decisions. Wedo so by defining an agent systemspecification that indicates possible decisions of agents. Weshowhowwe can derive agent decisions and preferencesfroman agent systemspecification. First we considerthe logic of rules we adopt. Logic of rules Thestarting point of any theory of decision is a distinction between choices madeby the decision makerand choices imposedon it by its environment. For example, a software upgrade agent (decision maker) mayhave the choice to upgrade a computersystem at a particular time of the day. The software company(the environmen0mayin turn allow/disallow such a upgrade at a particular time. Let S = {al,...,c~,} be the society or set of agents, then we therefore assumen disjoint sets of propositional atoms: A = AI U... U A, = {a, b, c,...} (agents’ decision variables (Lang1996) or controllable propositions (Boutilier 1994)) and W= {p, q, r,...} (the world parametersor controllable propositions). In the sequel we consider each decision makeras entities consisting of rules. Sucha decision makergenerates its decisions by applyingsubsets of rules to its input. Beforewe proceedsomenotations will be introduced. ¯ LA,, Lwand LAWfor the propositional languages built up fromthese atomsin the usual way, and variables z, y, ... to stand for any sentencesof these languages. ¯ CnA,, Cnwand GnAWfor the consequence sets, and ~A~, ~Wand ~aw for satisfiability, in any of these propositionallogics. ¯ z =~ y for an ordered pair of propositional sentences calleda rule. ¯ ER(T)for the R extension of T, as defined in Definition 1 below. In our frameworkthe generation of decisions are formalized basedon the notion of extension. In particular, the decision of an agent, whichis specified by a set of rules R and 38 Wegivesomeproperties oftheR extension ofT inDefinition 1. First note that ER(T)is not closed underlogical consequence. The following proposition showsthat ER(T) is the smallest superset of T closed underthe rules R interpreted as inferencerules. Proposition 1 Let ¯ E°(T) = ¯ E~(T) = EiR-I(T) U R(CnAw(EiR-l(T)))fori We have ER(T) = U~°E~(T). This definition of extensionsallowsinconsistent decisions as illustrated by the followingexample. Example 1 Let R = {T =~ p,a =~ ~p} and T = {a}, where T stands for any tautology like p V -~p. Wehave ER(O) = {p} and ER(T) = {a,p,-~p}, i.e. the R extension ofT is inconsistent. Of course, allowing inconsistent decisions maynot be intuitive. Weare here concernedabout possible decisions rather than reasonableor feasible decisions. Later wewill define reasonableor feasible decisions by excludinginconsistent decisions. Agent system specification Anagent systemspecification given in Definition 2 contains a set of agentsandfor eachagenta descriptionof its decision problem. The agent’s decision problemis defined in terms of its beliefs and desires, whichare consideredas belief and desire rules, a priority orderingon the desire rules, as well as a set of facts and an initial decision(or prior intentions). assumethat agents are autonomous,in the sense that there are no priorities betweendesires of distinct agents. Definition 2 (Agentsystem specification) An agent system specification is a tuple AS = (S, F, B, D, >_, °) t hat contains a set of agents S, and for each agent ~i a finite n F, set of facts Fi C Lw (F = [.Ji=l i), afinite set of belief rules Bi C LAW× Lw(B = uin=l Bi), a finite set of desire rules Di C_ LAWx LAW(D = Ui~l Di), a relation _>~_cD, x D~(>_=[.J~n__ 1 _>,) whichis a total ordering(i.e. reflexive, transitive, andantisymmetricandfor any two elements dl and d2 in Di, either da >_i d2 or d2 >_i dl), and a finite initial decision 6° C LA (6o = Uinl 6O). For an agent ai ¯ S we write x ~i y for one of its rules. In general, a belief rule is an orderedpair x =~4y with x E LAWand y E Lw. This belief rule should be interpreted as ’the agentai believesy in contextx’. Adesire rule is an ordered pair x =~i Y with x E LAWand y E LAW. Thisdesire rule shouldbe interpreted as ’the agent desires y in contextx’. It impliesthat the agent’sbeliefs are aboutthe world(x =~4P), and not about the agent’s decisions. These beliefs can be about the effects of decisions madeby the agent(a =~i P) as well as beliefs aboutthe effects of parameters set by the world(p =~4q). Moreover,the agent’s desires canbe aboutthe world(x =~i P, desire-to-be), but also about the agent’s decisions (x =~i a, desire-to-do). Thesedesires can be triggered by parametersset by the world (p ~i Y) well as by decisions madeby the agent (a =~i y). Modelling mentalattitudes such as beliefs and desires in terms of rules results in whatmightbe called conditional mentalattitudes (Broersenet al. 2001b). Agent Decisions The belief roles of the form LA ~ Lw can be used to determine the expected consequencesof a decision, where a decision54 of agent ai is any subsetof LA, that containsthe initial decision5i°. Theset of expectedconsequences of this decision 5i is the belief extensionof Fi U 5i. Moreover,we considera feasible decisionas a decision that doesnot imply a contradiction. Definition 3 (Decisions) Let AS = ({al .... ,a,},F,B,D,>,50) be an agent system specification. AnAS decision profile 5 for agents al ..... an is 5 = (51,... 5n) where5i is a decision of agent such that 50 C 5i C LA, fori = 1...n A feasible decisionfor agentai is 54 suchthat EB,( Fi U5i ) is consistent A feasible decisionprofile is a decisionprofile suchthat EB(Fu 5) is consistent Thefollowingexampleillustrates the decisionsof a single agent. Example2Let A1 = {a, b, c, d, e}, W = {p,q} and AS = ({al }, F, B, D, E, °) with F1 - -- { ~p}, B1 ={c q,d ~ q,e =~ ~q}, DI = {T =~ a,T =~ b,b ~ p,T =~ q,d =~ q}, >1= {b =*- p > T =~ b}, and5° = {a}. The initial decisiono51 reflects that the agenthasalreadydecided in an earlier stage to reachthe desire T =~a. Notethat the consequentsof all B1 rules are sentences of Lw, whereas the antecedentsof the B1rules as well as the antecedents and consequentsof the D1 rules are sentences of LAW.We havedue to the definition of ER(S): EB(FU {a}) = {--,p, a} EB(Fu {a,b}) = {--,p,a,b} Es(FU {a,c}) = {-~p,a,c,q} EB(FU{a,d}) = {--,p,a,d,q} EB(FU {a, e}) = {--,p, a, e, ~q} EB(FU{a,d,e}) = {-~p,a,d,e,q,-~q} 39 Therefore{a, d, e} is not a feasible ASdecision profile, becauseits belief extension is inconsistent. Continuedin Example5. According to definition3, the feasibility of a decisionprofile, whichindicates the decisions of individual agents, is formulatedin terms of the consistencyof the extension that is calculated basedon the decision profile. However,there are two different waysto calculate such an extension dependingon whetheragents can apply rules of each other or not. First, the extension can be calculated by demanding that each agent generatesits decisions by applyingonly its ownbelief and desire rules. This wayof calculating the extension assumesthat mental attitudes of agents are private and inaccessible to each other. Wewill call this notion of feasible decision profile individualistic. Second,the extension can be calculated by allowingagents to apply belief and desire rules of each other. This notion of feasible decision profile, whichwill be called collective, assumesthat mental attitudes of agents are public. Definition4 Feasibledecision profiles can be defined in the following two ways: ¯ An individualistic feasible decisionprofile 5 is a decision profile if EB( F 0 5) = ~i=l EBi(Fi U54) is consistent. ¯ A collective feasible decisionprofile5 is a decisionprofile if EB(FU S) = EB,u...uB, (F1 U 51 U . . . UUSn ) is consistent. Thefollowingexampleillustrates individualistic and collective feasible decisionprofiles. Example3 Let A1 = {a}, A2 = {b}, W = {p} and AS = ({al,a2},F,B,D,E,50 ) with F1 = F2 = 0, B1 = {a =~ p}, B2 = {b =~ -~p}, D1 = D2 = 0, >_ is the identity relation, and 5o = 5° = 0. The decision profile 5 = (51,52) = (0, {b}) is both individualistic and collective since EB~uB2(F1 0 51 0 F2 0 52) EBb(F1U 51) U EB2(F2U (I2) {b,-~p}. Bu t, th e decision profile 5 = (51,52) = ({a}, {b}) is neither individualistic nor collective since EB~UBz(FI U 51 U F2 U 52) EB, (z~’~l U 51) E,2(F2U 52) = {a,p, b, ~p}. Finally, the followingexampleillustrates that a decision profile canbe individualisticbut not collectivefeasible decision profile. Example 4 Let A1 = {a}, A2 = {b}, W= {p, t} and AS = ({al, a2}, F, B, D, >, 5°) with F1 = F2 = 0,B1 = {a =~ p}, B2 = {b ~ t,p ~ ~t}, D1 = D2 = 0, >_ is the identity relation, and60 =60 =0. Thedecision profile 5 = (51,52) = ({a}, {b}) is individualistic feasible since EBa(F1USI)UEB2(F2U52)= {a,p}U{b,t} = {a,p,b,t}. However, 5 is not a collective feasible decisionprofile since EBIuBz(F1U51 UF2 052) = {a,p,b,t,-~t}. In the following,we use the definition of individualistic feasible decisionprofiles. Agent preferences In this section weintroduce a wayto comparedecisions. We comparedecisions by comparingsets of desire rules that are not reachedby the decisions. Since an agent systemspecification providesonlythe orderingon individual desire rules, andnot orderingonsets of desire rules, wefirst lift the ordering on individual desire rules to an ordering on sets of desire rules. In this way, we makeit possible to compare sets of desire rules using the preferenceon individualdesire rules. Definition 5 Let AS = (S, F, B, D, >, 6O) be an agent systern specification,Di, ~Di" two subsets of Di, and Di \ i the set of Di elements whichare not D[ elements. Wehave D[ _D >-i ifVd "" " ED#i i\Di3d , i Di\DEi isuch thatd I > dII. Wewrite D[ >- D[’ if D[_>-Di" and Di" ~ Di,’ and we write D~~- D~’ if D~~_ D" and D~’~ D~. Thisdefinitionstates that an agent~ prefers a set of desire rules D~ to anotherset of desire rules D"if for eachdesire rule in D"there exists a preferable desire rule in D~. Thus, if D’ contains only one desire rule whichis preferable by c~ to any desire rule from D", then a prefers D~ to D". The followingpropositionsshowthat the priority relation ___is reflexive, anti-symmetric, and transitive. Proposition2 Let DI (£ D3and D39~ D1. For finite sets, the relation ~- is reflexive ( VDD >- D), anti-symmetric VD1, D2 (D1 ~- D2 ^ D2 ~- D1) ~ D1 = D2, and transitive, i.e. D1~- D2and D2>- Daimplies Dz ~- D3. The desire rules are used to comparethe decisions. The comparisonis based on the set of unreacheddesires and not on the set of violated or reached desires. A desire z =~y is unreachedby a decision if the expectedconsequencesof this decision implyx but not F- Thedesire rule is violated or reached if these consequencesimply x A --,y or x A y, respectively. Definition 6 (Comparingdecisions) Let AS = (S, F, B, D, >_, °) be a n agent s ystem specification and 6 be a AS decision. The unreacheddesires of decision 6for agent ai are: Decision 6 is at least as good as decision 6’ for agent ai, written as 6 >iv6’, iff u,(6’) _>u,(6) Decision 6 dominatesdecision 6~ for agent ai, written as 6 >6’, iff 6 6’ and 6’ 6 Thus,a decision6 is preferredby agent c~ to decision6~ if the set of unreached desire rules by 6 is less preferredby ~ to the set of unreacheddesire rules by 6~, i.e. the unreached desire rules by 6 are less importantthan unreacheddesire rules by 6’. The followingcontinuation of Example2 illustrates the comparisonof decisions. Example 5 (Continued) Wehave: U({a}) = {T ~ b,T ~ q}, U({a, b}) = {b =~ p, T =~ q}, U({a, c}) = {T =>. q}, 4O U({a, d}) = {T::~ b, d =~ q}, U({a, e}) = {T:=~ b, T :=~ q}, U({a,b, c})= {bp}. Wethus havefor examplethat the decision {a, c} dominates the initial decision {a}, i.e. {a, c} >u {a}. Thereare two decisions for whichtheir set of unreachedcontains only one desire. Dueto the priority relation, wehavethat {a, c} > u {a, b, c}. From Agent System Specifications to Game Specifications In this subsection, we consider agents interactions based on agent system specifications, their correspondingagent decisions, and the ordering on the decisions as explained in previous subsections. Gametheory is the usual tool to modelthe interaction betweenself-interested agents. Agents select optimal decisions under the assumptionthat other agents do likewise. This makesthe definition of an optimaldecision circular, and gametheory therefore restricts its attention to equilibria. For example,a decision is a Nash equilibrium if no agent can reach a better (local) decision by changing its own decision. The most used concepts from gametheory are Pareto efficient decisions, dominantdecisions and Nash decisions. Wefirst repeat somestandard notations from gametheory (Binmore1992; Osborne&Rubenstein 1994). As mentioned,we use 6i to denote a decision of agent t~ and 6 = {61,..., 6n) to denote a decision profile containing one decision for each agent. 6-i is the decision profile of all agents except the decision of agent o~i. (6_~, 6[) denotes a decision profile whichis the sameas ~ except that the decision of agent i from 6 is replacedwith the decision of agent i from6’. 6/ >iv 6i denotesthat decision~[ is better than 6i accordingto his preferences>iv and 6~>iv6i if better or equal.A is the set of all decisionprofiles for agents al,..., c~n, Af C A is the set of feasible decision profiles, and Ai is the set of possible decisionsfor agent ai. Definition 7 (GameSpecification) A gamespecification is a tuple (S, A, (>_i)) where ¯ S is a set (of agents) ¯ A = Az x ... x A,~ where Ai is a set of decisions of agent oq ¯ (>_i) is a partial orderon A (agent i’s preferences) Wenowdefine a mappingAS2ACfrom an agent specification to a gamespecification. Definition 8 (AS2AG)Let AS = = (S {al,..., t~n}, F, B, D, >_, 6O) be specification of agent systemin S, Ai be the set of ASfeasible decisions of agent o~i according to Definition 3, A = A~ x ... x A,, and > vi be the AS preferencerelation of agent ai defined on A accordingto definition 6. Then, the gamespecification of ASis the tuple (S, A, (>iv)). Wenowconsider different types of decision profiles which are similar to types of strategy profiles from game theory. Definition9 Let AS be the set of feasible decisionprofiles. A PSdecisionprofile 6 = (61,..., 6n) E Af is: Paretodecision if there is no 6’ = (5~,..., 6~) E Af for which6~ >u6ifor all agents ~i. stronglyParetodecisionif there is no U= (6’x, . . , 6~) E A f for which6[ >_iv6i for all agentsoq and63 >u gj for someagents ~j. dominantdecision if for all 6’ E AI and for every agent v ’ 6i) i it holds: (6~_i, 61) >_i (6_i, ’ i.e. a decisionis dominant if it yields a better payoff than any other decisions regardlessof whatthe other agentsdecide. Nashdecision if for all agents i it holds: (6-i, 6i) _>iv (6_i,6~i)for all i E A} It is a wellknown fact that Paretodecisionsexist (for finite games), whereasdominantdecisions do not have to exist. Thelatter is illustrated by the followingexample. Example6 Let ~1 and a2 be two agents, FI = F2 = 9, and initial decisions 6° = 6° = 9. They have the following beliefs en desires: Do, = {T =~p,T =~ q} >at= T :> p > T => q > T =~ -w > T => -~ p B,~2 = {b ::> q,-~b ~ -~q} FromGameSpecifications to Agent Specifications D~,= {T~ -,p, T ~ ~q} >a,= T=~. ~p> T =>. ~q> T=>q> T=>p Let AS be feasible decision profiles, EB be the outcomesof the decisions, and U(ti) be the set of unreached desires for agent~i. AI EB (a, b) {p, q} U6, Example8 Let A1 = {a} (oq cooperates), A2 = {b} (52 cooperates), and AS be an agent system specification with D1 = {T :=> ~aAb, T =~ b, T => ~(aA~b)}, D2= {T => a A -~b, T ~ a, T =¢,--~(--~a A b)}. Theonly Nashdecision is {-~a, -~b}, whereasboth agentswouldprefer {a, b}. Starting from an agent systemspecification, we can defive the gamespecification and in this gamespecification we can use standard techniques to for examplefind the Pareto decisions. However,the problemwith this approachis that the translation froman agent systemspecification to a game specification is computationallyexpensive. For example,a compactagent representation with only a few belief and desire rules maylead to a huge set of decisions if the number of decisionvariables is high. The mainchallenge of qualitative gametheory is therefore whether we can bypass the translation to gamespecification, and define properties directly on the agent system specification. For example,are there particular properties of agent systemspecification for whichwe can provethat there alwaysexists a dominantdecision for its correspondingderived gamespecification? A simple exampleis an agent system specification in whicheach agent has the samebelief and desire rules. U~: {T =4--,p, T =’;, -,q} {T =’,, --,p} {a,-~b) {p,--q} {T =~, q} {--p, q} {T => p} {T =:> ~q} (-~a,b) (--,a,-~b) {-~p,--,q} {T =:,’- p, T =~ q} 0 0 According to definition 5, for Ctl : U((a,b>) >U((a,-~b>) U( (-~a,b)) >U( (-~,-~b)) andfor oc2: U((-~a,-,b)) > u(( ~a,b)) > U((a,-~b)) > u((a,b)). Noneof these decision profiles are dominantdecisions, i.e. the agents specifications has no dominantsolution with respectto their unreached desires. The followingexampleillustrates a typical cooperation game. Example7 B1 = {a =¢ p, b => -~p A q}, D1 = {T => pAq} B2 = {c ::~ q, d =~ pA-~q} D2 = {T ~ pAq}. The agents have a commongoal p A q, whichthey can only reach by cooperation. Thefollowingexampleillustrates a qualitative version of the notorious prisoner’s dilemma,wherethe selfish behavior of individual autonomousagents leads to global bad decisions. 41 In the previoussection a mappingfromagent systemspecifications to standard gamespecifications has beendefined. In this section the questionis raised whetherthe notion of agent systemspecification is expressive enoughfor gamespecifications, that is, whetherfor each possiblegamespecification there is an agent systemspecification that can be mappedon it. Weprovethis propertyin the followingstandardway: 1. First, we define a mappingfrom gamespecifications to agent systemspecifications; 2. Second, we showthat the composite mappingfrom game specifications to agent systemspecifications and back to gamespecificationsis the identity relation. The second step showsthat if a gamespecification GSis mappedin step 1 on agent system specification AS, then this agent systemspecification AS is mappedon GS. Thus, it showsthat there is a mappingfromagent systemspecifications to gamespecifications for every gamespecification GS. Unlike the second step, the composite mappingfrom agent systemspecifications to gamespecifications and back to agent systemspecifications is not the identity relation. Thisis a direct consequence of the fact that there are distinct agent system specifications that are mappedon the same gamespecification. For example, agent system specifications in whichthe variable namesare uniformlysubstituted by new names. The mappingfrom gamespecifications to agent system specificationsconsists of the followingsteps: 1. Copythe set of agents; 2. Introducea set of decision variables andinitial decisions that will generatethe set of decisions; 3. Introduce a set of parameters, and for each agent sets of initial facts, beliefs, and desires, such that they generate the preferenceorder for each agent. Westart with the set of decision variables. Weintroduce for each decision a separate decision variable. Moreover, we assumeinitial decisionsthat see to it that the agent selects at least onedecision variable, andat mostone decision variable. Definition 10 (Decisions) Let A = {~1,..., ~,~} be a set of decisions (of someagent). Thenwe introduce the set decision variables A = {dl ..... d~} (for this agent) and the followinginitial decision(for this agent): dl V ... V dn i 5~ j =~-,(di A dj) Moreover,wedefine the sets of desires. Definition11 Considera set of decisions with a partial ordering on it (for an agent). For eachdecision 6, for which weintroducedthe decision variable d, we introducethe followingdesire: T=~ V d’A A -’,d’ 6’>~ 6~’~’ and z,~ The following proposition showsthat the mappingfrom gamespecifications to agent systemspecifications leads to the desiredidentity relation for the composite relation. Proposition 3 Let GSbe a gamespecification as in Definition Z Moreover,let AS be an agent system specification suchthat the decisionvariables,the initial decisionsandthe desires are as definedin Definition10and11, andlet the set of parameters,the sets of beliefs, the set observationsbe empty. Themappingfrom this agent system specification to gamespecification defined in Definition 8 mapsAS to GS. Proof. Assumeany particular GS. Construct the AS as above. Nowapply Definition 6. The unreached desires of decision6 for agentc~i are: Ui(~) = {x =~ y E D I EB(Fu~) ~ x EB(FUtJ) ~ y} The subset ordering on these sets of unreacheddesires reflects exactlythe original orderingon decisions. The followingtheoremfollows directly fromthe proposition. Concluding remarks Weare using a very simple approach, probablythe simplest approachpossible, basedon a twosorted propositional language. This simplicity is not a disadvantagebut an advantage, for the following reasons: it enables us to focus on essentials (comparehowGelfondand Lifschitz (Gelfond Lifschitz 1993) simple .,4 languagewhichfocussedplanning logics on essentials), it is easy to implementor generalize to classical formalisms, and it showswhyexactly we need morecomplexstuff such as causal theories and action languages. Our motivation comesfrom the analysis of rulebased agent architectures, whichhave recently been introduced. However,the results of this paper maybe relevant for a muchwider audience. For example, Dennett argues that automatedsystemscan be analyzed using conceptsfrom folk psychologylike beliefs and desires. Our workmaybe used in the formal foundationsof this ’intentional stance’ (Dennett 1987). In this paper we have defined a qualitative decision and gametheory in the spirit of classical decision and gametheory. The theory illustrates the micro-macrodichotomyby distinguishing the optimization problem from gametheoretic equilibria. Wealso link decision theory to gametheory by showinghowagent system specification can be mapped to gamespecification and back. Wethink that the methodof this paperis moreinteresting than its formalresults. Thedecision and gametheory are based on several ad hoc choices which need further investigation. For example,the desire rules are defeasible but the belief rules are not (the obvious extension leads to wishful thinking problemsas studied in (Thomason2000; Broersen et aL 2001a)). Wehope that further investigationsalongthis line brings the theories and tools used for individual agents and multi agent systems closer together. There are several topics for further research. The most interesting question is whetherbelief and desire rules are fundamental,or whetherthey in turn can be represented by someother construct. Othertopics for further research are the developmentof an incremental any-time algorithm to find optimal decisions, the developmentof computationally attractive fragmentsof the logic, and heuristics of the optimization problem. References TheoremI The set of agent system specifications with emptyset of parameters,beliefs rules and observations is expressive enoughfor gamespecifications. Proof. Followsfromconstruction in Proposition3. Theaboveconstruction raises the question whetherother sets of agent systemspecifications are completetoo, suchas for examplethe set of agent systemspecifications in which the initial decisionis an emptyset. Or the set of agent systemspecifications in whichthe set of desires contains only desire-to-be desires. Weleave these questions for further research. 42 Binmore, K. 1992. Fun and Games: A Text on Game Theory. Boutilier, C. 1994. Toward a logic for qualitative decision theory. In Proceedingsof the KR’94,75-86. Broersen, J.; Dastani, M.; ; and van der Torre, L. 2001a. Resolvingconflicts betweenbeliefs, obligations, intentions and desires. In Symbolicand Quantitative Approachesto Reasoningand Uncertainty. Proceedings of ECSQARU’01, LNCS2143, 568-579. Springer. Broersen, J.; Dastani, M.; Huang,Z.; Hulstijn, J.; and van der Torre, L. 200lb. The BOIDarchitecture: Conflicts betweenbeliefs, obligations, intentions and desires. In Proceedingsof he Fifth International Conferenceon AutonomousAgents (AA2001), 9-16. ACM Press. Cohen,P., and Levesque,H. 1990. Intention is choice with commitment. Artificial Intelligence 42:213-261. Dastani, M., and van der Torre, L. 2002.A classification of cognitive agents. In Proceedingsof the 24th AnnualMeeting of the CognitiveScienceSociety (Cogsci’02). Dennett, D. 1987. The intentional stance. Cambridge,MA: MITPress. Doyle, J., and Thomason,R. summer1999. Background to qualitative decision theory. AI magazine20(2). Gelfond, M., and Lifschitz, V. 1993. Representingaction and change by logic programs. JLP17:301-321. Lang,J. 1996.Conditionaldesires andutilities - an alternative approachto qualitative decision theory. In In Proceedings of the European Conferenceon Artificial Intelligence ( ECAI’96 ) 318-322. Osborne, M. J., and Rubenstein, A. 1994. A Course in GameTheory. Cambridge,Massachusetts: The MITPress. Rao, A., and Georgeff, M. 1991a. Modelingrational agents within a BDIarchitecture. In Proceedingsof the KR91. Rao, A. S., and Georgeff, M.P. 1991b. Deliberation and its role in the formationof intentions. In Proceedingsof the SeventhConference on Uncertaintyin Artificial Intelligence (UAI- 91). Savage,L. 1954. Thefoundationsof statistics. Simon, H. A. 1981. The Sciences of the Artificial. Cambridge, MA:MITPress, secondedition. Thomason,R. 2000. Desires and defaults: A framework for planningwith inferred goals. In ProceedingsKR2000, 702-713. von Neumann,J., and Morgenstem,O. 1944. Theory of Gamesand EconomicBehavior. Princeton, NJ: Princeton UniversityPress, 1 edition. 43