Two-choice probability learning: Test of a finite markov chain and linear operator model

Abstract

Human subjects were trained in the noncontingent binary choice situation under one of two levels of event probability (π = .6, .8) and payoff (0c, 1c). These data were used to evaluate the ability of the Reinforcement-Extinction (R-E) Model and the Weak-Strong (W-S) Conditioning Model to describe marginal and first-order response-event dependencies throughout learning and second-order response-event dependencies and run curves during the last 100 trials. Both models described asymptotic statistics about equally well with largest discrepancies being in the fit-to-run curves. Generally, poorer preasymptotic fits were attributed to too rapid a rate of approach of the predicted statistics to their asymptotic values. Parameter estimates for both models varied with π and payoff throughout learning. From these estimates it was inferred that learning occurs on correct and incorrect trials; correct trials appeared more important in the R-E model while incorrect trials assumed major importance for the W-S model, especially prior to overshooting.