Learning Flexible Goal-Directed Behavior

Abstract:

I will present a general framework for skill learning that can be applied to many tasks including spatial navigation, sensory-motor learning, eye-movement control and cognitive task execution. Based on data from animal learning theory and child development it will be proposed that many goal directed behaviors can be explained as a combination of three mechanisms: (1) stimulus-approach, where available sensory data directly guide the action to the goal, (2) stimulus-response, where sensory data selects a response that leads toward the goal but the response is not controlled by the goal, and (3) contextual inhibition, where stimulus-approach or stimulus-response must be suppressed because those behaviors would not lead to goal-fulfillment in the particular situation. Together with the appropriate learning mechanisms, the framework explains the gradual differentiation and increased context sensitivity of behaviors and suggests how a child or a robot could learn an ever increasing behavior repertoire from interaction with the environment.