Two Learning Models for Two-Choice Conditioning Experiments Involving Nonreinforced Trials

Abstract

Four groups of 24 Ss each were run for 480 trials to investigate the effects of nonreinforced trials in a two-choice contingent verbal conditioning experiment. Two hypotheses concerning the effects on response probability of nonreinforced trials were considered. The two hypotheses, (1) an identity hypothesis, which amounted to assuming no change in response probability on nonreinforced trials, and (2) a sharing hypothesis, which amounted to assuming a change estimated from the data, were expressed in a linear model and a two-state Markov model, respectively. With appropriate selection of reinforcement parameters, the two models yielded different predictions for asymptotic response probability. On the basis of observed asymptotic response proportions, the linear-identity model was found more acceptable. Further tests indicated that the linear-identity model predicted pre-asymptotic response proportions rather well but failed to account adequately for variability of the responses.