Learning Experiments of a FALCON Using SVR for a Card Game

Abstract

This paper proposes a method to improve the learning performance of a learning agent with FALCON, to make a player agent for the card game “Hearts”, which is one of the multi-player imperfect-information games. FALCON is a machine learning method which is an extended fuzzy ART（Adaptive Resonance Theory）. The previous work showed that FALCON is effective for Hearts. In this study, to improve the learning performance, the action set of the agent is changed based on strategies of Hearts, and a method that employs a prediction by the support vector regression is proposed.