Evaluating case-based decision theory: Predicting empirical patterns of human classification learning

We introduce a computer program which calculates an agent’s optimal behavior according to Casebased Decision Theory (Gilboa and Schmeidler 1995) and use it to test CBDT against a benchmark set of problems from the psychological literature on human classification learning (Shepard, Hovland, and Jenkins 1961). This allows us to evaluate the efficacy of CBDT as an account of human decision-making on this set of problems. We find: (1) The choice behavior of this program (and therefore Case-based Decision Theory) correctly predicts the empirically observed relative difficulty of problems in the benchmark human data, which is a strong vote of confidence in its favor. (2) ‘Similarity’ (how CBDT decision makers extrapolate from memory) is decreasing in Euclidean vector distance, consistent with evidence in psychology (Shepard 1987). (3) Average similarity is rejected in favor of additive similarity. (4) CBDT learns the correct solutions unrealistically fast relative to human learners.