Authors:
Hitomi Morishita
;
Hiroaki Ueda
and
Kenichi Takahashi
Affiliation:
Hiroshima City University, Japan
Keyword(s):
Multi-agent Reinforcement Learning, FALCON, Multi-channel ART Networks, Hearts.
Related
Ontology
Subjects/Areas/Topics:
Agent Models and Architectures
;
Agents
;
Artificial Intelligence
;
Artificial Intelligence and Decision Support Systems
;
Computational Intelligence
;
Distributed and Mobile Software Systems
;
Enterprise Information Systems
;
Evolutionary Computing
;
Knowledge Discovery and Information Retrieval
;
Knowledge Engineering and Ontology Development
;
Knowledge-Based Systems
;
Machine Learning
;
Multi-Agent Systems
;
Soft Computing
;
Software Engineering
;
Symbolic Systems
Abstract:
3-channel fuzzy ART network FALCON is a good solution to combine reinforcement learning with state segmentation, where it learns the relations among percepts, actions and rewards. FALCON, however, does not have a mechanism to predict behavior of other agents, and thus it is difficult for FALCON to learn the optimal agent’s behavior in a multi-agent circumstance. In this paper, an action prediction module based on 2-channel fuzzy ART network is proposed, and FALCON is modified in order to be able to register the output of the action prediction module. The modified FALCON is called FALCON AP. Moreover, FALCON ER that estimates the expected value of rewards and selects an action according to the value is proposed. Through experiments in which FALCON, FALCON AP and FALCON ER are applied to a card game Hearts, it is shown that FALCON ER receives less penalty points and learns better rules.