Implementation of chaining in operant conditioning by a neural network circuit

Abstract

Chaining training is an effective training method in operant conditioning to help agents learn complex tasks that must occur in a specific sequential order. This approach deconstructs a complex task into its constituent components and systematically teaches each component in a step-by-step manner. This paper presents a memristive neural network circuit to implement chaining training in operant conditioning. The whole circuit is constructed from several single-behavioral training circuits connected in sequence, with the action output of one circuit serving as the cue input of the subsequent one. The single-behavioral training circuit consists of time-delay modules, a reward adjustment module, and a read/write circuit for synapse weight. This single-behavioral training circuit not only simulates basic processes in operant conditioning, but also demonstrates that the training speed continues to decrease due to reward fatigue. By introducing multiple rewards, the impact of reward fatigue can be mitigated. Chaining training is also successfully implemented for a two-step target task. Finally, this neural network circuit is applied to assembly robots, enabling them to perform grasping and installation tasks adaptively. This work has significant application potential in the field of industrial robotics.