A computer model of probabilistic concept formation

Abstract

The purpose of this investigation was the development and assessment of a revised model of probabilistic concept formation. The original model, CODER, was designed to form concepts based on the occurrence of covarying characteristics in the environment. The revised model, HAREM, extends the notion of covarying characteristics to conditions which include knowledge of results, conditions under which concept. formation may be regarded as guided as opposed to unguided. In addition, the separation of the attribute and rule learning processes, as suggested in the more traditional concept formation literature, is incorporated in HAREM. The revised model, HAREM, specified in the form of a computer program, was initially used in an effort to reproduce human performance in an unconstrained sorting task, this task required each S to sort VARGUS 7 stimuli from three schema families into as many categories as he desired. HAREM was then altered to be compatible with a sequential prediction paradigm. In this second task, the Ss were to make predictions as to what digit would occur next. Digit sequences used in this task were constructed from two separate five digit sequences. Individual sequences in the experiment exhibited only a probabilistic correspondence to these generation sequences (i.e., the digit sequences had the same probabilistic properties as VARGUS 7 stimuli). The value which provided the best fit to the human data for the parameter which controlled the encoding of attributes was lower than has previously been used in models of schematic concept formation. On the other hand, the value of this parameter in the sequential prediction task should perhaps have been closer to that used in previous models of schematic concept formation. It was inferred that the rate at which probabilistic concept formation takes place depends on the degree of structure of the task (i.e., the degree to which the relevant dimensions are discernable, as a function of the task, instructions, and presence or absence of knowledge of results). The simulations of HAREM for the two experimental tasks were only moderately successful. Discrepancies between the humans' and simulations' performance provided opportunities for speculation about means of improving the predictive utility of HAREM. Although no firm statement can be made, the results at least suggested that CODER's basic components can be utilized in constructing models applicable to several different varieties of probabilistic concept formation.