Association for Behavior Analysis International

More traditional computer simulations of human behavior have involved information processing models of the brain, whereby the brain is assumed to be modeled after the way computers are architected: lots of data stored in memory with programs that retrieve the data given certain inputs. Adaptive network systems are elemental computer learning programs that have enabled the simulation of behavior at an operant level, whereby that behavior which is followed by reinforcing consequences is more likely to occur again. This symposium explores how adaptive networks can generate simple verbal operants, and how more diverse and complex behavior can then be generated as a result. Implications for the interpretation of more complex human linguistic behavior and the development of effective teaching programs also are considered.

Psychologists, behavior analysts, and graduate students interested in exploring how adaptive networks can generate simple verbal operants, and how more diverse and complex behavior can then be generated as a result.

Adaptive networks are in essence, computer programs that learn. This very fact places them squarely in the domain of behavior analysis, although few in the field conduct research with them, or develop them for commercial or other purposes. This talk will provide an overview of what Adaptive Network Systems (ANS) are and how they differ from typical computer programs; provide insight into how they work; and show how they can adhere to the behavioral principles that the experimental analysis of behavior has uncovered over the years. The basic components of such a network will be explored including the architecture, some simple learning algorithms, and design features which preclude hard-wiring responses, or using brute computer power to solve problems or to produce more complex behaviors. Finally, some simple examples of ANS will be illustrated, particularly related to the elementary verbal operants.

Because the experimental analysis of verbal behavior is constrained by practical and ethical considerations, most of our understanding of complex cases arises from verbal interpretations. But such interpretations are limited by the sheer number of relevant variables and our ignorance of subjects' histories. In contrast, adaptive network simulations permit complete control over both complex contextual variables and historical variables. If such simulations are tightly constrained by behavioral principles, they offer powerful demonstrations of the explanatory adequacy of such principles. Dr. Palmer will discuss several examples that seem to defy verbal interpretation, examples such as the problems of novelty, nesting, generalization of neologisms according to apparent grammatical form, conditioning the behavior of the listener, mysterious structural regularities in verbal behavior, and the problem of acquisition of complex forms. He will suggest that adaptive network simulations of verbal behavior may be the best interpretive tool and in some cases the only one.

The presentation will describe a robot whose behavior is learned via an adaptive network based on behavior analytic principles, embedded in a body with sensors including vision and hearing and with responses including spatial movements and vocalizations. The demonstration will first show how that robot learns elementary verbal operants, then more complex verbal behaviors based on them. We will examine in detail how some of the puzzling verbal behaviors described in the preceding papers in the symposium are learned, illustrating how using a robot makes it possible to examine moment-to-moment changes in the conditions that control the behavioral sequence.