INCORPORATING PRIOR BIASES IN NETWORK MODELS OF CONCEPTUAL RULE LEARNING

A series of simulations is reported in which extant formal categorizat ion models are applied to human rule-learning data (Salatas & Bourne, 1974). These data show that there are clear differences in the ease wi th which humans learn rules, with the conjunctive the easiest and the biconditional the hardest. The original ALCOVE model (an exemplar-base d model), a configural-cue model, and two-layer backpropagation models did not fit the rule-learning data. ALCOVE successfully fit the data, however, when prior biases observed in human rule learning were imple mented into weights of the network. Thus, current empirical learning m odels may not fare well in situations in which learners enter the conc ept-formation situation with preconceived biases regarding the kinds o f concepts that are possible, but such biases might nevertheless be ca ptured within these models. By incorporating preexperimental biases, A LCOVE may hold promise as a comprehensive category-learning model.