发帖
雷达卡
[size=10.000000pt]Prescriptive Bayesian decision making has reached a high level of maturity, sup-ported by efficient, theoretically well-founded algorithms. However experimentaldata shows that real decision makers choose such Bayes-optimal decisions surpris-ingly infrequently, often making decisions that are badly sub-optimal. So preva-lent is such imperfect decision-making (versions of which are sometimes knownas “bounded rationality”) that it should be accepted as an inherent feature of realdecision makers living within interacting societies.
[size=10.000000pt]To date such societies have been investigated from an economic and game-theoretic perspective, and even to a degree from a physics perspective. However,little work has been done from the perspective of computer science and associ-ated disciplines like machine learning, information theory and neuroscience. Theworkshop [size=10.000000pt]Decision Making with Multiple Imperfect Decision Makers[size=10.000000pt], held after the24th Annual Conference on Neural Information Processing Systems (NIPS 2010) inWhistler, British Columbia, Canada, was a step in bringing such alternative view-points to bear on the topic of understanding societies of imperfect decision makers.This book includes extended versions of selected contributions from the workshop.
[size=10.000000pt]A prescriptive (normative) theory of perfect decision-making can serve as a goldstandard against which we can compare systems of interacting imperfect decision-makers. Doing so suggests that many societal, biological, and engineered systemsof interacting imperfect decision makers make their joint decisions in a reasonablysuccessful manner. Analyzing the structure, rules and behaviour of such systems isthe central theme of this book.
[size=10.000000pt]Some of the questions that arise in such an analysis include:
- How should we formalise rational decision making of a single imperfect decisionmaker?
- Does the answer change for a system of imperfect decision makers?
- Can we extend (modify) existing prescriptive theories for perfect decision makers
to make them useful for imperfect ones?
[size=9.000000pt]VI
[size=9.000000pt]Preface
[size=10.000000pt]••
[size=10.000000pt]How can we exploit the relation of these problems to the control under varyingand uncertain resources constraints as well as to the problem of the computationaldecision making?
What can we learn from natural, engineered, and social systems to help usaddress these issues?
[size=10.000000pt]The chapters in this book address some of these questions, add others, and hope-fully will stimulate new ones. Some of the particular topics they focus on are
- the meaning of rationality in a multiple participant setting,
- combination of knowledge and preferences,
- judicious use of information,
- hybrid (human and machine) decision making systems,
- scalability of negotiation processes,
- the relationship between prescriptive and descriptive decision making,
- how decision making is done in living organisms.
The particular contributions are as follows:
A. Carlin and S. Zilberstein inspect meta-reasoning that allows imperfect decisionmakers acting in decentralised setting to stop their computations at an appropriatetime while contributing to optimization of an overall time-dependent utility. Theyprovide a novel treatment of anytime algorithms in a setting of multiple participants.
M. Ka ́rny ́ and T.V. Guy describe a proper extension of Bayesian decision making,that offers a single imperfect decision maker a methodology for sharing knowledgeand preferences. It serves for efficient selfish decision making within a multipleparticipants’ setting.
J. Rubin, O. Shamir and N. Tishby consider the interaction of multiple imperfectdecision-makers within the framework of Markov decision processes. By consid-ering the information flow between decision makers they arrive at a novel designtechnique and at randomized decision strategies. Their approach allows them to es-tablish a proper tradeoff between a decision’s value and the processed information.
R. Lee and D.H. Wolpert combine Bayesian networks and game theory to derivea framework for predicting and controlling a system containing both human andautomated decision makers. They elaborate their framework for predicting aircraftpilot behaviour in potential near mid-air collisions.
K. Fujita, T. Ito and M. Klein deal with negotiation protocols, a key ingredient dis-tinguishing single and multiple decision makers’ settings. They address scalabilitylimitations inherent to problems with imperfect decision makers by proposing issue-grouping. This allows them to go beyond traditional negotiation mechanisms thatrely on linear utilities and improve the efficiency of negotiation outcomes.
Y.H. Chang, R. Levinboim and R. Maheswaran address the discrepancy between pre-scriptive (predicted) and descriptive (observed) behaviour of decision makers in theultimatum game. They model this discrepancy by accounting for how participants
[size=9.000000pt]Preface VII
[size=10.000000pt]are inevitably influenced by the past actions taken. They benchmark their modelwith observations made on real players.
[size=10.000000pt]A.E.P. Villa, P. Missonier and A. Lintas [size=10.000000pt]investigate the relationship of living neu-ronal systems with associated decision making activities. Working with rats, theirfirst study tries to find experimentally how the decision making effort maps on mea-surable signals. Their second study inspects EEGs of people playing the ultimatumgame. Their experiments bring new insight into physiology of decision making andindicate the extent of difficulties in relating of brain activities and decision making.
[size=10.000000pt]Acknowledgements. [size=10.000000pt]The Institute of Information Theory and Automation, Academyof Sciences of the Czech Republic, NASA Ames Research Center, the Center forNonlinear Studies, and the Santa Fe Institute all supported us in preparing this book.The editors from Prague were also supported by GAC[size=10.000000pt]ˇ [size=10.000000pt]R 102/08/0567.
[size=10.000000pt]July 2011 Tatiana V. GuyPrague, Moffett Field Miroslav Ka ́rny ́David H. Wolpert
京公网安备 11010802022788号
