MICROSTRUCTURE OF GUESS PROCESSES: PART C

Abstract : Trial-to-trial changes in the proportion of human subjects predicting the occurrence of one of two events in a complex sequence of binary events (probability learning) are analyzed in terms of several simple models. The direction of change predicted by linear-operator reinforcement models (Estes, Bush and Mosteller) is wrong on about 75% of the trials. A no-learning model, a time-dependent decay model, and a cycle-dependent decay model are used to provide some insight into the nature of probability learning. Some suboptimal procedures for estimating parameters of stochastic processes are compared. The method of minimum absolute error is recommended as being very useful.