Learning to Detour

Abstract

Anurans (frogs and toads) show quite flexible behavior when confronted with stationary objects on their way to prey or when escaping from a threat. Rana computatrix (Arbib, 1987), an evolving computer model of anuran visuomotor coordination, models complex behaviors such as detouring around a stationary barrier to get to prey on the basis of an understanding of anuran prey and barrier recognition, depth perception, and appropriate motor pattern generation mechanisms based on sensory perception. Our present analysis of detour behavior goes beyond other models by incorporating new data from our laboratory demonstrating a learning component in anuran detour behavior. Building on earlier work showing how interacting schemas may be used to analyze a complex environment to generate an appropriate course of behavior, we turn to the questions: How are the relevant schemas adapted? How are schemas combined to form new schema assemblages acquired for the system to become more efficient? We describe the construction mechanisms and interactions with the environment that are necessary to achieve higher levels of detour performance. We have based this article mostly on data about learning to detour when approaching prey, but the model offers a strategy for learning to detour in general. Moreover, we have attempted to solve the problem in a general way so that the model of learning to detour points the way to a general theory of schema-based learning.