Open-Ended Inference of Relational Representations in the COSPAL Perception-Action Architecture

Abstract

The COSPAL architecture for autonomous artifical cognition utilises incremental perception-action learning in order to generate hierarchically-grounded abstract representations of an agent's environment on the basis of its action capabilities. We here give an overview of the top-level relational module of this architecture. The first stage of the process hence involves the application of ILP to attempted action outcomes in order to determine the set of generalised rule protocols governing actions within the agent's environment (initially defined via an a priori low-level representation). In the second stage, imposing certain constraints on legitimate first-order logic induction permits a compact reparameterisation of the percept space such that novel perceptual-capabilities are always correlated with novel action capabilites. We thereby define a meaningful empirical criterion for perceptual inference. Novel perceptual capabilities are of a higher abstract order than the a priori environment representation, allowing more sophisticated exploratory action to be taken. Gathering of further exploratory data for rule induction hence takes place in an iterative cycle. Application of this mechanism within a simulated shape-sorter puzzle environment indicates that this approach significantly accelerates learning of the correct environment model.