Judith Spencer Reitman

Cognitive systems based on adaptive algorithms

The type of cognitive system (CS) studied here has four basic parts: (1) a set of interacting elementary productions, called classifiers, (2) a performance algorithm that directs the action of the system in the environment, (3) a simple learning algorithm that keeps a record of each classifiers success in bringing about rewards, and (4) a more complex learning algorithm, called the genetic algorithm, that modifies the set of classifiers so that variants of good classifiers persist and new, potentially better ones are created in a provably efficient manner.

Skilled perception in Go: Deducing memory structures from inter-response times☆

Experts appear able to handle much larger amounts of specialized information than nonexperts, and handle it without an apparent superior memory capacity. This finding, based on research on chess players with chess information, was replicated on Go players with Go information. Assuming this superiority occurs because the experts process chunks of information through their limited capacities rather than individual elements, the question then becomes one of defining what the chunks are and how they are related. To this end, the technique of partitioning recall and reproduction data into chunks on the basis of inter-response times (IRTs) (introduced in their work on chess by Chase and Simon, 1973) was applied to the reproduction and recall of Go patterns by a Go Master and a Go beginner. Unlike its application in chess, no single IRT was able to produce consistent, veridical chunks for either Go player. Subsequent analysis of the underlying assumptions of the technique showed it to be limited to only those patterns that can be partitioned into a linear set of chunks, not nested chunks, and to situations in which retrieval and overt recall of each chunk is completed before retrieval of the next chunk. In a supplementary task, the Master Go player indicated that the Go patterns were not seen as linear chunks nor as strictly nested hierarchies, but rather as overlapping clusters. IRTs were found to be correlated with this structure, but were not reliable enough to reflect its details.

Without surreptitious rehearsal, information in short-term memory decay

Reitman (1971) found that subjects could retain three words perfectly for 15 sec while detecting tones in noise and supposedly avoiding rehearsal. These results were taken to indicate lack of support for the decay principle of STM. Two studies reported here test two assumptions in the Reitman study: that 100% recall reflects not a ceiling effect but the absence of forgetting, and the lack of disruption of interpolated detection performance indicates lack of rehearsal. Major results indicated that (1) the 1971 study did involve a ceiling effect; (2) tonal detection is measurably disrupted when subjects rehearse; and (3) when subjects detect equally well in the retention interval as in a control interval they forget 33% of what they can recall immediately, and when they detect syllables instead of tones, they forget about 44% more. There is clear evidence for both decay and simple interference in STM.

Storage and later recognition of exemplars of concepts

This paper concerns the problem of abstraction: whether when we encounter several exemplars of a concept, we retain only the abstracted concept, only the exemplars, or both. Although many studies concur that both are stored, a recent article argued strongly that only the abstracted concept is stored. The present study, aimed at replication of this recent finding, follows the earlier procedural details but adds appropriate controls and uses simpler material. A set of 24 exemplars of four concepts, in the form of four-tuples of letters and numbers, was presented to Ss who, after presentation, rated a larger set of exemplars for recognition. One group of Ss experienced the conceptual exemplars; control group Ss experienced items that were similar in composition but not exemplars of a concept. Two major results appeared: Unlike the study on which this was based, all Ss were able to distinguish those items that were originally experienced from those that were not. And, the more completely an exemplar fit the concept (the longer the item), the more confident the S was that it had been presented. In contrast, in the control condition, the longer the item, the more confident the S was that it had not been presented earlier. Two models are described to account for these results. One is based on the Ss initial storage of the exemplars in a concept-plus-correction format; the other is based on a procedure whereby the S can make recognition judgments without having previously abstracted and retained the concepts.

COGNITIVE SYSTEMS BASED ON ADAPTIVE ALGORITHMS1

The type of cognitive system (CS) studied here has four basic parts: (1) a set of interacting elementary productions, called lassifiers, (2) a performance algorithm that directs the action of the system in the environment, (3) a simple learning algorithm that keeps a record of each classifiers success in bringing about rewards, and (4) a more complex learning algorithm, called the genetic algorithm, that modifies the set of classifiers so that variants of good classifiers persist and new, potentially better ones are created in a provably efficient manner. Two “proof-of-principle” experiments are reported. One experiment shows CSs performance in a maze when it has only the ability to adjust the predictions about ensuing rewards of classifiers (similar to adjusting the “weight” of each classifier) vs. when the power of the genetic algorithm is added. Criterion was achieved an order of magnitude more rapidly when the genetic algorithm was operative. A second experiment examines transfer of learning. Placed in a more difficult maze, CS with experience in the simpler maze reaches criterion an order of magnitude more rapidly than CS without prior experience.

Mnemonic elaboration in multilist learning

This study inquires whether retroactive interference (RI) from learning multiple word lists can be altered through mnemonic strategies. The S s learned five successive lists of 20 words each by associating the words with 20 conceptual pegs via visual imagery. Some S s were instructed to visualize the words from the successive lists in entirely new associative scenes; other S s were to incorporate the current lists words into the appropriate scenes from the earlier lists. Although equivalent in immediate recall, the Separate Images (SI) group recalled less than the Progressive Elaboration (PE) group at the end of the session, and showed a strict RI curve across lists where the PE group did not. Recall at a one-week test, however, favored earlier lists for the PE group only, explicable in terms of cumulative rehearsal of earlier items with the PE method.

Goals and plans in a program for playing Go

A program that plays Go provides a basis for analyzing possibilities for extending present AI conceptions of planning and goal structures to problems that are ill-structured, dynamic, multiperson, resource-bound, and highly interactive. The focus is on mechanisms for communicating information and control over time and among a number of interacting processes, in a flexible but coherent manner. Using the capabilities of current computer languages, it is possible to specify planning and goal structures, and appropriate conventions for them, that will accommodate the demands of these increasingly complex and sophisticated problem environments.