Being and acting rational

Being and acting rational
of 3
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  Being andActing Rational IEEEINTERNETCOMPUTING1089-7801/03/$17.00©2003 IEEEPublished by the IEEEComputer SocietyMARCH • APRIL 2003 91  Agents on the Web Michael N.Huhns • University of South Carolina • I usually prefer to deal with rational people. Inthe vernacular, being rational  means beingreasonable — that is, using reason when mak-ing decisions, taking actions, and achieving goals.Rational people tend to be more predictable and,therefore, understandable. Understanding a deci-sion makes it easier for me to accept it. When I encounter and interact with agents onthe Web, or when my own agents encounter them,I prefer that these agents behave rationally as well. An agent can behave rationally in at least threeways. Some of these might be better than others;some might be appropriate for individuals but notfor groups. Rationality Types  A rationality theory indicates what is rational andwhat is not in specific cases. Three such theoriesthat govern an agent’s behavior are logical ratio-nality  , economic rationality  , and  pragmatic ratio-nality  . They depend, respectively, on the mathe-matics of logic, probability, and computation. Toconform to their corresponding mathematical for-malism, each type requires strong assumptionsabout the nature of a rational agent’s world andhow the agent can sense and act in that world. Logical Rationality In the fourth century BC, the Greek philosopher  Aristotle formulated a process for reasoning thatwould lead to irrefutable conclusions. His systemof syllogisms (a kind of inference mechanism) couldproduce not only new knowledge, but also a meansto achieve goals based on logically justified actions.The resultant view — that reasoning could bespecified precisely and thus mechanized — evolvedin the early 20th century into the doctrine of logi-cal positivism , which held that everything an agentknows can be derived from observation sentencesthat represent the agent’s environment. The deriva-tions come from applying laws of deductive logic.Deductive logic provides rational constraints onbelief in two ways. First, it can be used to definethe notion of deductive consistency and inconsis-tency: deductive inconsistency determines a kindof incoherence in belief. Second, the laws of de-ductive logic can constrain admissible changes inbelief by providing deductive rules of inference  . For example, modus ponens is a deductive rule of inference that requires Q to be inferred from sen-tences P  and P  → Q .So, to be logically rational, an agent “simply”has to convert everything it senses into a sentence(a belief) in a formal language, combine the sen-tences with all other sentences it has sensed or derived, derive new sentences about its world, anduse this new set of sentences to choose its actions.Several major problems cloud this approach. First,observations about the world might be uncertain andincomplete, making them difficult to express logi-cally. Second, several courses of action could lead toa goal’s achievement, and it is difficult for logic tohelp an agent decide among them. Third, there mightnot be any action that an agent can prove willachieve its goal, leaving the agent without help indeciding what to do. Finally, reasoning about a largeset of sentences might be intractable. Economic Rationality  Another option is for agents to be economically rational. 1 Like logically rational agents, economi-cally rational ones act to achieve their goals on thebasis of what they know. Operationally, however,an economically rational agent ranks possibleactions by the expected utility of their results andthen executes the action that has the highestexpected utility. (Expected utility is defined interms of the agent’s possible actions, the probabil-ities of the actions’ outcomes, and the agent’sranked preferences among those outcomes.) Putsimply, economic rationality is based on decisiontheory, which combines logic and probability the-ory with utility theory to provide a means for mak-ing decisions under uncertainty.   Applying probability theory to ratio-nality is attributed to the ReverendThomas Bayes (c. 1701–61). 2 Bayesianepistemology’s two main features arethe introduction of a  formal apparatus for inductive logic and the introductionof a  pragmatic self-defeat test  for epis-temic rationality as a way to extend justification of the laws of deductivelogic to include inductive logic. 3 Theformal apparatus adds standards of probabilistic coherence and a rule of probabilistic inference, both of whichapply to degrees of belief (degrees of confidence). Bayesian decision theory is now the dominant theoretical modelfor both the descriptive and normativeanalyses of decisions. Unfortunately, economic rationali-ty requires a computationally expen-sive search over the outcomes of allpossible sequences of actions (becauseseveral actions might be needed toachieve a goal), knowledge of theprobability distributions for the out-comes (which are difficult to deter-mine), and a means for assigning util-ities to outcomes. 4 Pragmatic Rationality The earlier approaches to rationality rely on the assumption that the worldwill not significantly change whilethe agent decides what to do, and thatan action that is rational when deci-sion-making begins will be rationalwhen it concludes. Clearly, this isproblematic in real-world settings.Imagine a car-driving agent waitingat a stop sign. The agent looks bothways, does not see any other vehicles,and then, remembering that itdropped its map, rummages on thefloor for several minutes to find it.Finally, with map in hand, it deducesfrom its observations that there are noapproaching vehicles and drivesblithely across the intersection.Similarly, if proving that it wassafe to cross the intersection took longer than the time for the trafficstatus to change, the resultant proof would be worthless. A pragmatic ap-proach takes such computational lim-itations into account.  According to Stuart Russell and Peter Norvig, 2 this means doing “the rightthing.” Formally, they define this as “For each possible percept sequence, a ratio-nal agent should select an action that isexpected to maximize its performancemeasure, given the evidence provided by the percept sequence and whatever built-in knowledge the agent has.” When agents have limited abilities,the best they can achieve is boundedrationality. The ideal of logical or eco-nomic rationality requires choosingactions to maximize a measure of expected utility. That utility shouldreflect a complete and consistent pref-erence order and probability measureover all possible contingencies. Thisrequirement appears too strong to per-mit an accurate specification for realis-tic individual agent behavior. We can weaken the ideal require-ments for rationality in many ways.Possibilities include anytime algorithmsthat return the best action found eachtime they terminate and theories thatattempt to mimic human decision mak-ing. Because of the rich variety of psy-chological types we can observe inhumans — each with different reason-ing strengths and limitations — it isunlikely that there will be a single bestapproach to pragmatic rationality. 5 Rationalityand Multiagents Rationality is important for groups of agents as well. If all agents in a groupare individually rational, no matter what type of rationality they use, willthe group necessarily behave rational-ly? That is, will the group always makethe decisions, take the actions, andachieve the goals that are best for it?This question is important for govern-ments, political organizations, corpo-rations, teams, and committees. An agent’s best strategy oftendepends on what strategies other a - gents choose. For each agent in agroup to behave rationally by maxi-mizing its self-interest, for example, itmust consider the behaviors of other agents who are also behaving in their own self-interest. This consideration isthe basis of game theory, which pro- vides mathematical guidance for howagents in a multiagent system decidetheir actions.In open or continuous environ-ments, deciding what is best dependson a time horizon — it is usually impractical for agents to reason infi-nitely far into the future or to consid-er an infinite number of intermediatestates. For a given finite time horizon,an agent must choose a strategy thatconsiders either the consequences of  just the end result or the consequencesof both the ends and the means. Whenthe agents are part of a society, ethicscan provide some guidance. 92MARCH • APRIL 2003  Agents on the Web Figure 1.Rationality types and their major characteristics.Formal approachesto rationality,such as logical and economic rationality,have severe limitationswhen scaled up and applied in real-world domains.Theories of bounded ratio-nality are more closely related to human reasoning and appear more promising for practical implementations of agents. Ad hocRationalityLogicalEconomicPragmaticSmall domainsUnsound and/or incompleteActions constrainedby decision theoryPropositionallogicFirst-orderlogicNonclassical logic(temporal, modal)BoundedrationalityActions constrainedby deductionActions constrained bytheories of computation  The ethical theory egoism holds that actionshould maximize self-interest. A parallel theory called utilitarianism holds that action shouldmaximize the universal good of all agents. Boththeories consider only the end result (they are tele-ological  ); they hold that the best thing to do isalways maximize a certain good, in which goodcan be interpreted as pleasure, preference satis-faction, interest satisfaction, or aesthetic ideals. Incontrast, deontological  theories hold that the endsdo not justify the means, and agents must at eachstep choose the action that does not endanger society’s welfare. Conclusion Rationality alone is insufficient to specify agentdesign. Using economic theory, we can programagents to behave in ways that maximize their utili-ty while responding to environmental changes.However, economic models for agents, althoughgeneral in principle, are typically limited in practicebecause the value functions that are tractable essen-tially reduce an agent to acting selfishly. 2 Buildinga stable social system from a collection of agentsmotivated by self-serving interests is difficult. Finally, understanding rationality and knowl-edge requires interdisciplinary results from artifi-cial intelligence, distributed computing, econom-ics and game theory, linguistics, philosophy, andpsychology. A complete theory involves semanticmodels for knowledge, belief, action, and uncer-tainty; bounded rationality and resource-boundedreasoning; commonsense epistemic reasoning; rea-soning about mental states; belief revision; andinteractions in multiagent systems. Acknowledgment The US National Science Foundation supported this work under grant number IIS-0083362. References 1.M.N. Huhns and L.M. Stephens, “Multiagent Systems andSocieties of Agents,” Multiagent Systems , G. Weiss, ed., MITPress, 1999, pp. 106–111.2.S. Russell and P. Norvig,  Artificial Intelligence: A Modern Approach, 2nd ed. , Pearson Education, 2003.3.W. Talbott, “Bayesian Epistemology,” The Stanford Ency-clopedia of Philosophy  ;, Reasoning About Rational Agents, MITPress, 2000.5.J. Doyle, “Rational Decision Making,” MIT Encyclopedia of Cognitive Science  ; Michael N. Huhns is a professor of computer science and engi-neering at the University of South Carolina, where he alsodirects the Center for Information Technology. IEEEINTERNETCOMPUTING     A    D    V    E    R    T    I    S    E    R     /     P    R    O    D    U    C    T     I    N    D    E    X MARCH / APRIL 2003 John Wiley &SonsBack Cover Advertising PersonnelAdvertiser / Product Page Number Marion Delaney IEEEMedia, Advertising DirectorPhone:+1 212 419 7766Fax:+1 212 419 7589Email: Marian Anderson Advertising CoordinatorPhone:+1 714 821 8380Fax:+1 714 821 4010Email: Sandy Brown IEEE Computer Society,Business Development ManagerPhone:+1 714 821 8380Fax:+1 714 821 4010Email: Debbie Sims Assistant Advertising CoordinatorPhone:+1 714 821 8380Fax:+1 714 821 4010Email: Advertising Sales Representatives Mid Atlantic (product/recruitment) Dawn BeckerPhone: +1 732 772 0160Fax: +1 732 772 0161Email: Midwest (product) Dave JonesPhone: +1 708 442 5633Fax: +1 708 442 7620Email: dj.ieeemedia@ieee.orgWill HamiltonPhone: +1 269 381 2156Fax: +1 269 381 2556Email: wh.ieeemedia@ieee.orgJoe DiNardoPhone: +1 440 248 2456Fax: +1 440 248 2594Email: Connecticut (product) Stan GreenfieldPhone:+1 203 938 2418Fax:+1 203 938 3211Email: Northwest (product) John GibbsPhone: +1 415 929 7619Fax: +1 415 577 5198Email: Midwest (product) David KovacsPhone: +1 847 705 6867Fax: +1 847 705 6878Email: Midwest/Southwest recruitment) Tom WilcoxenPhone: +1 847 498 4520Fax: +1 847 498 5911Email: New England (recruitment) Barbara LynchPhone: +1 401 739 7798Fax: +1 401 739 7970Email: Southern CA (recruitment) Tim MattesonPhone: +1 310 836 4064Fax: +1 310 836 4067Email: Japan German TajiriPhone:+81 42 501 9551Fax:+81 42 501 9552Email: Europe (product) Hilary TurnbullPhone:+44 131 660 6605Fax:+44 131 660 Europe (recruitment) Penny LeePhone:+44 020 7405 7577Fax:+44 020 7405 7506Email:Pennylee@Essentialmedia.Co.Uk Southeast (product/recruitment) C. William Bentz IIIEmail:  bb. ieeemedia@ieee.orgGregory MaddockEmail:  gm. ieeemedia@ieee.orgSarah K. WileyEmail:  sh. ieeemedia@ieee.orgPhone:+1 404 256 3800Fax:+1 404 255 7942 Southwest (product) Royce HousePhone: +1 713 668 1007Fax: +1 713 668 1176Email: Southern CA (product) Marshall RubinPhone: +1 818 888 2407Fax: +1 818 888 4907Email: New England (product) Jody EstabrookPhone: +1 978 244 0192Fax: +1 978 244 0103Email:
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks