Transfer of response in visual recognition situations as a function of frequency variables.

A series of 6 experiments investigated the principles required to account quantitatively for responses of human Ss to new combinations of cues following discrimination learning. In the training phase of each experiment, Ss learned to make identifying responses (numerical labels) to sets of stimuli (pairs of nonsense syllables) under a paired-associate procedure. After a fixed number of acquisition trials Ss were tested on stimulus compounds involving new combinations of the training cues. In all experiments a substantial proportion of variance in the test data was accounted for by a model embodying the additive rule and the probability matching rule of stimulus sampling theory. In cases when training had been conducted under standard discrimination paradigms, responses to test compounds were quite well accounted for by this model without auxiliary principles. Ss exhibited preferences for low ambiguity cues in test compounds only when this preference had been differentially reinforced during training. Under a variety of circumstances, predictions from the stimulus sampling model could be improved by the addition of a 'relative novelty' principle, stating that, other things being equal, Ss tend to sample from test compounds the cues that had occurred least frequently during previous training.