The Bactra Review: Occasional and eclectic book reviews by Cosma Shalizi   94

Modeling Bounded Rationality

by Ariel Rubinstein

MIT Press, 1998

O Docta Simplicitas!

"Bounded rationality" is a term of the economists' art, introduced (apparently) by Herbert Simon in the 1950s. The standard-issue Homo economicus is un-boundedly rational, i.e. the decisions he reaches are always the best ones which could be made, given the evidence available to him (and he always correctly hedges against his ignorance and uncertainty). Now, this is simply not the way people are (as is obvious, if not when we examine our own decisions, then certainly those of our acquaintances). There are also real and apparently insuperable computational difficulties which make it most doubtful that any creature could actually solve the necessary optimization problems, at least within a reasonable time. So people must be boundedly rational --- rational within certain limits; muddling through on some tricks (academically: procedures) which usually don't serve us too shabbily. Nor do we optimize: we satisfice.

So much (here and now) for the idea of bounded rationality according to Simon. In this vague form, it simply must be true. The problem is that, while there is only one course of action open to an ideal optimizer (namely, the optimum), there are about as many outcomes for procedural satisficers as there are satisficing procedures. (There are many ways to bound rationality.) Worse, the mathematical methods for modeling substantially rational agents are highly developed, but we have very little grip on boundedly rational agents of any description. This what's called the "can't beat something with nothing" justification for continuing to make models with unboundedly rational agents.

Rubinstein's book aims to provide another "something," i.e. a set of tools for modeling boundedly rational agents of various sorts. These are: simplifying decision-making by ignoring "similar" features of incompatible options; limiting knowledge (represented by a formalism borrowed from modal logic); limiting memory; aggregating individual decisions by teams; and playing games by means of automata (i.e. limiting game-theoretic strategies to those which can be implemented by abstract automata). Some of these sections --- I am thinking particularly of those on limited memory, on teams, and on automata-games --- are highly suggestive, and probably very useful to workers in robotics and artificial agents. In all cases, the math is developed very elegantly.

With admirable generosity, Rubinstein includes a critique of this work by Herbert Simon. Uncle Herbert's message is, in effect: these are all very pretty, now let's find out which have psychological validity, so we can use them for real economics; and, by the way, the real thing probably won't be nearly so tidy as any of these, which is why we need cognitive science and AI. Rubinstein's reply is equally forthright. "Models of economic theory are meant to establish `linkages' between the concepts and statements that appear in our daily thinking on economic situations" (p. 191), not predict or explain facts about economic life, or even get those facts right, or even offer bench-marks on what is possible. In other words, economics is not an empirical science, but an analytical refinement and vindication of common sense. Rubinstein does say that he'd like his models to have some kind of contact with reality, but thinks this is underwritten by starting from everyday, intuitive notions. To be perfectly honest, I am flabbergasted that Rubinstein can at once elaborate his models with such care and subtlety, and be so cavalier about whether they the basic ideas they contain are any more true than those in newspaper editorials.

Of course, there is no need to follow Rubinstein on this point to profit from the rest of his book. Doing so does need a good command of standard tools of microeconomics (expected-utility-maximizing decision theory, game theory), and some acquaintance with formal logic would probably be helpful. Given those, it will be very useful to economists, to other social scientists to a much lesser degree, and to computer programmers interested in designing multi-agent systems.

xii + 208 pp., bibliography, index of names and subjects (good), some black-and-white diagrams
Artificial Life and Agents / Economics
Currently in print as a hardback, ISBN 0-262-18187-8 [buy from Powell's], US$29.95, and as a paperback, ISBN 0-262-68100-5 [buy from Powell's], US$16.95, or as a PDF from Prof. Rubinstein's webpage, free. LoC HD30.23 R83 1998.
30 September 1999