Advances in Decision-Theoretic AI:
Limited Rationality and Abstract Search

Abstract:

This thesis reports on the current state of ongoing research by the author and others investigating the use of decision theory---its principles, methods, and philosophy---in Artificial Intelligence. Most of the research discussed in this thesis concerns decision problems arising within the context of Computer Game-Playing domains, although we discuss applications in other areas such as Clinical Decision-Making and Planning as well. We discuss in detail several AI techniques that use decision-theoretic methods, offer some initial experimental tests of their underlying hypotheses, and present work on some new decision-theoretic techniques under development by the author.

We also discuss some attempts, by the author and others, to transcend a few of the limitations of the earlier approaches, and of traditional decision theory, and synthesize broader, more powerful theories of abstract reasoning and limited rationality that can be applied to solve decision problems more effectively. These attempts have not yet achieved conclusive success. We discuss the most promising of the current avenues of research, and propose some possible new directions for future work.

Availability and Redistribution Policy

Please do not redistribute paper or electronic copies of this thesis yet, because it is currently undergoing major revision in preparation for a TR version. However, the web URL may be freely redistributed, and anyone may obtain copies of the thesis directly through this web page, which will always present the most up-to-date version or state that the thesis is temporarily unavailable, in which case copies may be requested via email to the author, mpf@ai.mit.edu.

Source (FrameMaker), Postscript (165 pages, 872K), Compressed PS (312K)

Contents: (Postscript)

Title page, Abstract, Acknowledgments (6 pages, 30K)
Table of Contents (4 pages, 37K)
Body chapters:

Introduction (6 pages, 38K)
Evolution of Computer Game Playing Techniques (88 pages, 400K)
Limited Rationality (32 pages, 368K)
Abstract State-Space Search (12 pages, 66K)
Summary and Conclusions (4 pages, 30K)

Bibliography (13 pages, 68K)

Citation information:

Michael P. Frank. Advances in decision-theoretic AI: Limited rationality and abstract search. Master's thesis, Massachusetts Institute of Technology, Cambridge, Massachusetts, May 1994. Available on the World-Wide Web at URL http://www.ai.mit.edu/~mpf/papers/Frank-94/Frank-94.html.

@MASTERSTHESIS{Frank-94,
        AUTHOR = {Michael P. Frank},
        TITLE = {Advances in Decision-Theoretic {AI}: Limited Rationality
                 and Abstract Search},
        SCHOOL = {Massachusetts Institute of Technology},
        YEAR = 1994,
        ADDRESS = {Cambridge, Massachusetts},
        MONTH = may,
        note = {Available on the World-Wide Web at URL {\tt
                http://www.ai.mit.edu/~mpf/papers/Frank-94/Frank-94.html}},
        url = {http://www.ai.mit.edu/~mpf/papers/Frank-94/Frank-94.html},
}