Irwin D. Nahinsky

A hypothesis sampling model for conjunctive concept identification

Abstract A Markov chain model for conjunctive concept identification in a two-category task is presented. Its basic assumptions involve sampling a subset of hypotheses compatible with an initial positive instance and eliminating untenable hypotheses with incoming information. Predicted values are compared with observed values for a number of statistics including error-run data and hypothesis-sample states for various trials. The model was found to fit fairly well.

The Self-Ideal Correlation as a Measure of Generalized Self Satisfaction

Each member of four groups, neurotics, inpatient psychotics, outpatient psychotics, and general population controls, Q sorted 100 items to describe himself and his ideal self. The general population control group showed a significantly higher average self-ideal correlation than did the diagnostic groups. This superiority was maintained even after specific items accounting for the intergroup differences were removed. It was shown that most of the between groups variance for the self-ideal correlation could be attributed to a generalized set transcending specific traits.

Statistics and moments-parameter estimates for a duoprocess paired-associate learning model

Abstract A model postulating all-or-none unlearning of wrong alternatives as well as acquisition of correct alternatives is presented. Moments estimates based on the total error distribution are developed for c , the learning rate constant, and for a , the unlearning rate constant. Statistics for testing the model are presented. They are: mean and variance for total errors, theoretical learning curve, trials until last error, and distribution of responses until the first trial upon which some alternative occurs for the last time.

A test of axioms of all-or-none concept identification models

Three groups of 24 S s each were presented respectively three concept identification tasks varied with respect to degree of overlap between instances and non-instances of the concept. Specific concepts and presentation orders were balanced with tasks in a 3 × 3 Greco-Latin square design. Examination of error and hypothesis-guessing data indicated that S s permanently eliminated many hypotheses early and did not sample them subsequently. Data also indicated that S s processed information on correct R trials.

Hypothesis sampling in conjunctive concept identification

The experiment entailed two 2 by 2 Latin square designs, each of which involved two conjunctive concept attainment problems balanced with respect to presentation order over two groups. Three blank trials followed an initial instance for each problem, and each blank trial stimulus overlapped the instance by exactly two attribute values. Analysis of patterns of blank trial Rs indicated that Ss sampled hypotheses about combination rather than single attribute values independently of each other.

All-or-none elimination of errors in paired-associate learning

A group of 18 Ss was presented a paired-associate learning task, with 8 nonsense syllables as stimuli, and the numbers 1 through 8 as possible Rs for each stimulus. The Ss were presented the stimuli successively in random cycles. A duoprocess (DP) all-or-none model showed better fits to the data than did a uniprocess (UP) all-or-none model for learning-rate constant, number of correct Rs for non-learned association trials, and proportion of correct Rs to total Rs on non-learned association trials. Stationarity of correct R probability was found to hold for trials until the first trial upon which some wrong alternative occurred for the last time and was found not to hold after this point until the last error. This accorded with DP assumptions. The predicted DP error curve fit the data adequately, but the predicted UP curve departed significantly from the data.

Language patterns in conjunctive concept identification

Subjects learned two conjunctive concepts utilizing geometric figures in which there were two relevant attributes. An initial positive instance in each problem was followed by three blank-trial stimuli which overlapped the initial instance by exactly two attribute values. The analysis of blank-trial responses indicated that Ss sample adjective-noun hypotheses significantly more often than adjective-adjective hypotheses, where adjective-noun and adjective-adjective describe labels which can be given the cue pairs.

The effect of number of attribute-values in conjunctive concept identification

A factorial experiment in conjunctive concept identification was performed in which number of attributes and number of attribute values were varied. Only number of attributes had a significant effect upon hypothesis sampling behavior, a finding which supports the notion that Ss sample only hypotheses compatible with positive instances as they are presented.

A test of two all-or-none models for paired-associate learning

A 4 × 4 factorial experiment was performed using a self paced paired-associate learning task. The independent-variable dimensions were number of stimuli and number of response alternatives, with each variable assuming values 2, 4, 6, and 8. Predictions for two all-or-none paired-associate learning models were compared with the data. One model predicted all-or-none learning of correct alternatives with no effect for incorrect responses, while the other model predicted all-or-none learning of correct alternatives and all-or-none elimination of incorrect alternatives. Results indicated that the former model was more appropriate in the two-response alternatives case while the latter model was more appropriate for four and six alternatives. Neither model fitted the eight-response alternative data because of the failure to account for a forgetting process.

Conjunctive hypothesis sampling: A reconsideration

Ss were presented with conjunctive concept learning problems in which blank-trial stimuli were used to assess hypothesis sampling patterns. Data were analyzed to show that Ss did not sample single values randomly and independently and respond on the basis of these single values. Evidence was also presented to show that Ss do not sample hypotheses with replacement after an error.