Faster learning in embodied systems through characteristic attitudes

Abstract

Classical reinforcement learning is a general learning paradigm with wide applicability in many problem domains. Where embodied agents are concerned, however, it is unable to take advantage of the structured, regular nature of the physical world to maximise learning efficiency. Here, using a model of a three joint robot arm, we show initial learning accelerated by an order of magnitude using simple constraints to produce characteristic attitudes, implemented as part of the learning algorithm. We point out possible parallels with constraints on the movement of natural organisms owing to their detailed mechanical structure. The work forms part of our EMBER framework for reinforcement learning in embodied agents introduced and developed in 2004.