The Ghost in the Machine: Inferring Machine-Based Strategies from Observed Behavior

We introduce a procedure to infer the repeated-game strategies that generate actions in experimental choice data. We apply the technique to set of experiments where human subjects play a repeated Prisoner's Dilemma. The technique suggests that two types of strategies underly the data.

Rational Choice on Arbitrary Domains: A Comprehensive Treatment

The rationalizability of a choice function on arbitrary domains by means of a transitive relation has been analyzed thoroughly in the literature. Moreover, characterizations of various versions of consistent rationalizability have appeared in recent contributions. However, not much seems to be known when the coherence property of quasi-transitivity or that of P-acyclicity is imposed on a rationalization. The purpose of this paper is to fill this significant gap. We provide characterizations of all forms of rationalizability involving quasi-transitive or P-acyclical rationalizations on arbitrary domains.

Domain Closedness Conditions and Rational Choice

The rationalizability of a choice function on an arbitrary domain under various coherence properties has received a considerable amount of attention both in the long-established and in the recent literature. Because domain closedness conditions play an important role in much of rational choice theory, we examine the consequences of these requirements on the logical relationships among different versions of rationalizability. It turns out that closedness under intersection does not lead to any results differing from those obtained on arbitrary domains. In contrast, closedness under union allows us to prove an additional implication.

Cooperative of Noncooperative Behavior?

In an abstract two-agent model, we show that every deterministic joint choice function compatible with the hypothesis that agents act noncooperatively is also compatible with the hypothesis that they act cooperatively. the converse is false.

Toward a rational-choice foundation of non-additive theories

A classical argument of de Finetti holds that Rationality implies Subjective Expected Utility (SEU). In contrast, the Knightian distinction between Risk and Ambiguity suggests that a rational decision maker would obey the SEU paradigm when the information available is in some sense good, and would depart from it when the information available is not good. Unlike de Finetti's, however, this view does not rely on a formal argument. In this paper, we study the set of all information structures that might be availabe to a decision maker, and show that they are of two types: those compatible with SEU theory and those for which SEU theory must fail. We also show that the former correspond to "good" information, while the latter correspond to information that is not good. Thus, our results provide a formalization of the distinction between Risk and Ambiguity. As a consequence of our main theorem (Theorem 2, Section 8), behavior not-conforming to SEU theory is bound to emerge in the presence of Ambiguity. We give two examples of situations of Ambiguity. One concerns the uncertainty on the class of measure zero events, the other is a variation on Ellberg's three-color urn experiment. We also briefly link our results to two other strands of literature: the study of ambiguous events and the problem of unforeseen contingencies. We conclude the paper by re-considering de Finetti's argument in light of our findings.