SUPERSTITION IN ARTIFICIAL NEURAl NETWORKS: A CASE STUDY FOR SELECTIONIST APPROACHES TO REINFORCEMENT

Abstract

The superstition phenomenon remains a crossroad of conceptual issues, especially regarding the operant-respondent distinction and the role of neural principies in understanding of behavior. In the present paper, I examine the phenomenon from the perspective of artificial neural networks, in the context of a selectionist approach to reinforcement. I define the basic phenomenon as a persisting change in a behavior that is not a conditional part of the reinforcement operation. Two computer simulations of this phenomenon were run using two feedforward and fully-connected selection networks. Superstition was obtained in both networks through the same reinforcement mechanism used to obtain Pavlovian and operant conditioning in previous simulations.

Results showed that response-dependent reinforcement was not necessary to change any emitted behavior, and that superstition was maximally generalized over the networks' repertoire. A more specific form of superstition was obtained in a third simulation by using a partially-connected network. A similar result might be obtained by making different responses mutually exclusive through inhibitory connections. Also, it is likely that a form of shaping through response-dependent reinforcement will be required in order to simulate more complex environment-behavior relations in selection networks. I conclude by examining certain criticisms that have been raised towards neural-network modeling in behavior analysis and the incorporation of neural principies in our accounts of behavior.