Abstract. Machine rule induction was examined on a difficult categorization problem by applying a Hollandstyle classifier system to a complex letter recognition task. A set of 20,000 unique letter images was generated by randomly distorting pixel images of the 26 uppercase letters from 20 different commercial fonts. The parent fonts represented a full range of character types including script, italic, serif, and Gothic. The features of each of the 20,000 characters were summarized in terms of 16 primitive numerical attributes. Our research focused on machine induction techniques for generating IF-THEN classifiers in which the IF part was a list of values for each of the 16 attributes and the THEN part was the correct category, i.e., one of the 26 letters of the alphabet. We examined the effects of different procedures for encoding attributes, deriving new rules, and apportioning credit among the rules. Binary and Gray-code attribute encodings that required exact matches for rule activation were compared with integer representations that employed fuzzy matching for rule activation. Random and genetic methods for rule creation were compared with instance-based generalization. The strength/specificity method for credit apportionment was compared with a procedure we call "accuracy/utility." Keywords. Category learning, parallel rule-based systems, exemplar-based induction, apportionment of credit, fuzzy-match rule activation.Human experts often solve difficult problems quickly and effortlessly by categorizing complex situations as special cases of familiar paradigms and applying solution strategies that are known to be effective for these paradigms (de Groot, 1965; Chase & Simon, 1973). Problem solving in this context involves partitioning a complex task into two components that can be solved independently and executed in a serial fashion. The first component consists primarily of categorization. Humans acquire this ability after many years of observing a wide-range of related examples. The expert's skill seems to be based primarily on memory for past experiences rather than on logical deduction or symbolic reasoning (Charness, 1981). The second component involves associating one or more action sequences with each of the categories. The problem solver has a large repertoire of well-practiced action routines that can be selected and applied in a way that is appropriate for the initial categorization decision.Our research focuses on the first component of the above paradigm. We examine a computer system that induces general categorization rules within a supervised learning paradigm. A large number of unique examples are presented to the system along with an outcome

No abstract

A formal conditioning model is proposed that adds a dynamic attention rule and a novel response mapping rule to the Rescorla-Wagner associative axiom. This model retains the virtues of its predecessor and, in addition, accurately simulates many conditioning phenomena that are not encompassed by the original Rescorla-Wagner model: (a) The acquisition function is S-shaped, and abrupt shifts in responsiveness occur during acquisition and extinction; (b) reacquisition is characteristically more rapid than the initial acquisition; (c) the behavioral effect of a set of conditioning operations is not necessarily independent of the subject's prior conditioning history. In addition, the new model copes better than the Rescorla-Wagner model with latent inhibition and with the ineffectiveness of nonreinforcing a conditioned inhibitor.The Rescorla-Wagner conditioning model differs from the traditional linear operator model proposed by Estes and Burke (19S3) and by Bush and Hosteller (19SS) in assuming that all stimuli share a common pool of associative strength so that their aggregate values, rather than their individual values, approach an overall asymptote. When one stimulus gains associative strength, the unconditional stimulus (US) becomes less effective in supporting conditioning to other stimuli that are presented in conjunction with the original stimulus. This novel assumption permits a successful extension of the linear model to account for overshadow-

A series of three experiments replicated and extended earlier research reported by Chase and Simon (1973), de Groot (1965), and Charness (Note 1). The first experiment demonstrated that the relationship between memory for chess positions and chess skill varies directly with the amount of chess-specific information in the stimulus display. The second experiment employed tachistoscopic displays to incrementally "build" tournament chess positions by meaningful or nonmeaningful chunks and demonstrated that meaningful piece groupings during presentation markedly enhance subsequent recall performance. The third experiment tested memory for one of two positions presented in immediate sequence and demonstrated that explanations based on a limited-capacity short-term memory (Chase & Simon, 1973) are not adequate for explaining performance on this memory task.The game of chess provides a useful working environment for the analysis of specialized informationprocessing skills. Skill in chess is acquired only with extensive exposure to the game. From an analysis of verbal protocols, de Groot (1965) established that chess masters and less able players use similar thought processes in analyzing a complex chess position. They consider a similar number of moves (about 35), calculate to similar depths (about 7 plies), make the same number of fresh starts (about 7), and analyze a similar number of moves per minute (about 3). The only major difference de Groot noted was that the masters invariably analyzed stronger moves than the weaker players. This conclusion (i.e., better players are better because they select better moves) was not terribly illuminating.de Groot's (1965) research did indicate that masters differed from weaker players in their ability to recall a chess position from an unfamiliar game after it had been presented for only 5 sec. Masters recalled 93% of 22 pieces, while strong club players recalled about 51%. Chase and Simon (1973) have shown that this recall ability is chess specific by replicating de Groot's result and including a control condition in which the chess pieces were arranged randomly. With a 5-sec exposure of quiet middle-game positions, their master recalled 81% of the 22 pieces, while a novice recalled only 33%. With the randomized positions, all subjects recalled only three to five pieces correctly. This result indicates that the master and novice have similar visual memory capacities for nonmeaningful piece configurations. Thus, the ability to play chess well seems to depend on aThe design and interpretation of these experiments benefited greatly from discussions with Neil Charness, Eliot Hearst, Henry Helff, Michael Humphreys, and Benton J. Underwood. The authors thank Stanton Tripodis for data collection assistance in Experiment I. Reprint requests should be addressed to Peter Frey, Department of Psychology, Northwestern University, Evanston, Illinois 60201. learned perceptual skill rather than on the acquisition of a sophisticated problem-solvingstrategy. EXPERIMENT IOur first study attemp...

Students' ratings of their instructors in undergraduate classes in calculus were correlated with class performance on a common final examination. Ratings on several instructional factors were highly related to class performance even though they appeared to be independent of the students' own grades.