A Reinforcement Learning Agent with Associative Perception

Z Zatuchna, AJ Bagnall

Research output: Contribution to conferencePaper

Abstract

One of the most perspective ideas of further development of Reinforcement Learning (RL) research involves using associative learning models to improve performance of reinforcement learning agents. Learning Classifier Systems (LCS) have proved to be one of the most successful classes of RL methods that have been applied to maze environments. However, so far LCS have shown their effectiveness for small sized and simple maze environment tasks only. We try to overcome the limits by tying up the connection between LCS performance and principles of established psychological phenomena, those of associative learning in particular. We bring together the ideas of imprinting, laws of organization and stimulus generalization to create a basis for introducing an associative perception and recognition to the LCS framework. As a result, we develop the Associative Perception Learning Model, a new concept for modelling the learning process in autonomous learning agents. The model has been implemented as AgentP, a new LCS with Associative Perception and its performance has been evaluated on existing and new maze problems.
Original languageEnglish
Pages92-99
Number of pages8
Publication statusPublished - 2006
EventSymposium on Associative Learning and Reinforcement Learning at Adaptation in Artificial and Biological Systems (AISB'06) -
Duration: 1 Jan 2006 → …

Conference

ConferenceSymposium on Associative Learning and Reinforcement Learning at Adaptation in Artificial and Biological Systems (AISB'06)
Period1/01/06 → …

Cite this