James R. Erickson

Probabilistic discrimination learning with dimensionalized stimuli

Abstract Two groups of subjects served in a probabilistic discrimination experiment where the stimuli contained four binary dimensions and the individual stimuli varied in II values. One group of subjects could use either individual cues (dimensions) which varied in their validity, combinations of cues, or the entire stimulus pattern as the basis of responding, but the other could only use the entire stimulus pattern. The performance of the first group was superior. More detailed examination of the data and fits of several different types of models to the data support the conclusion that subjects in the first group used primarily the single most valid dimension or perhaps the combination of the two most valid dimensions as the basis for responding. An expectancy model presented by Castellan and Edgell fit the data from both groups quite well.

Error elimination in paired-associate learning

All-or-none models seem to account for some aspects of paired-associate data fairly well, yet nonstationarity of performance during precriterion trials is often observed, a finding not consistent with these models. This paper considers the possibility that an increase in guessing efficiency due to elimination of errors accounts for nonstationarity. A modified pairedassociates procedure allowing for a complete ranking of all response alternatives to each stimulus on every trial was used. Several statistics were developed which provide strong evidence for the existence of error elimination. Reexamination of the precriterion trials proved that nonstationarity was not removed when corrections were made for an error-elimination process.

Concept identification with reinforcement of hypotheses and misinformative stimulus labeling

College students solved simple concept-identification problems in which feedback was contingent on their current hypothesis (H). No dimension was consistently labeled until Ss pattern of responses over a block of stimuli indicated that he was using the solution H. The results indicated that some Ss were pure H testers, choosing new Hs on the basis of local consistency; these Ss had little difficulty with their problems. The remaining Ss evidently tried to integrate the unreliable feedback information over more than one trial and found their problems quite difficult or impossible to solve.

Patterns of event runs in probability learning

Forty-eight Ss were assigned to a 2 by 2 factorial design with two levels of risk (1¢ and A¢) and two levels of sequence (patterned-run distribution and random-run distribution) in a two-choice probability learning task. All Ss received 300 trials. Conditional predictions were not affected by differences in risk. A significant effect of the sequence variable was found in responding conditional on runs of the less frequent event, but not in responding conditional on runs of the more frequent event. It was concluded that Ss respond to certain higher order regularities in the event-run structures.

Effects of a memory aid in probability learning

Eighty Ss served in a probability-learning experiment where π (.70 or.90), instructions (random or problem solving), and memory aid (present or absent) were varied factorially. Memory aid consisted of lights which informed S which events had been correct on the three previous trials. Conditional response probabilities were closer to conditional event probabilities in memory-aid conditions, especially under problem-solving instructions.

Reduction of the guessing pool in paired-associate learning

The Ss learned two 12-item paired-associate lists under self-paced conditions where the response pool was available. Analysis focused on S’s precriterion responses. Responses learned to items tended not to be made as guesses on other items, but this type of response-pool reduction was not sufficient to account for nonstationarity of the precriterion learning curve. Other sources of response-pool reduction such as partial learning of the correct response or stimulus generalization errors seem capable of accounting for the observed amount of nonstationarity.

Some characteristics of hypothesis sampling in concent identification

Two studies were conducted in which Ss solved a concept-identification problem whose solution was a hypothesis (H) recently tried, but which had been incorrect when it was tried. In both experiments, the difficulty of the final problem was linearly related to the recency with which the solution H had previously been sampled. When the solution was an H that S had not as yet tried, problem difficulty was approximately the same as when the solution was the H three back from the start of the final problem. Analysis of the H-sampling data revealed that Ss made only partial use of information from previous stimuli when choosing a new H to test. A tentative model of H sampling is presented which takes account of these and other data on use of memory while solving concept-identification problems.