Learning Action Affordances and Action Schemas

Abstract

Several theories of action selection assume that the environment both suggests and constrains possible actions at each moment. We present an interactive activation model (based on an existing model of routine sequential action selection) in which actions are organised into partially ordered schemas for simple task elements, and the current state of the environment contributes to selecting actions and schemas. Previous versions of the model have not accounted for learning. We show that a simple reinforcement learning paradigm allows environment-action associations to be acquired through unguided exploration of the environment. The basic model is limited in the types of environment/action associations that it acquires. We explore ways in which greater diversity of learning (and behaviour) may be achieved, and suggest that, in order to acquire a broad range of environment/action associations, mechanisms for boredom avoidance or novelty seeking are required.