Escape, Avoidance, and Imitation: A Neural Network Approach

Abstract

We present a real-time neural network that integrates classical and operant processes to describe how animals learn to escape and avoid an aversive stimulus either by trial and error or by imitation. It is assumed that through classical conditioning animals build an internal model of the environment and that through operant conditioning animals select from alternative responses. The internal model of the environment provides predictions of the aversive stimulus based on environmental stimuli and the animal's own responses, and these predictions are used to train the operant conditioning block to generate responses that minimize the aversive stimulus. Computer simulations show that the model correctly describes many of the features that characterize escape and avoidance. The network also is able to describe the imitation of a demonstrator by an observer. During the demonstration, a neural network representing the observer stores classical associations between environmental stimuli and the demonstrator's responses and aversive stimuli, and these associations serve to train the operant associations during the observer's performance. It is assumed that the demonstrator's responses evoke a representation of identical responses in the observer and that the demonstrator's unconditioned response to the aversive stimulus serves as an aversive reinforcer for the observer. The network contributes to a general theory of adaptive behavior and is relevant to the design of autonomous systems that learn either through trial and error or through imitation.