Improvement of Agent Learning for a Card Game Based on Multi-channel ART Networks
Kenta Nimoto, Keníchi Takahashi, Michimasa Inaba · Journal of Computers · 2016
The 3-channel fuzzy adaptive resonance theory network FALCON (Fusion Architecture for Learning, COgnition, and Navigation) is recognized as an effective method for combining reinforcement learning with state segmentation, in which learning targets the relationships between percepts, actions, and rewards.It has been shown that FALCON is effective in making a playing agent for the card game Hearts, although the agent was unable to beat a rule-based agent.This study proposes new learning methods for training FALCON, in order to create a stronger agent.Using percepts chosen with an emphasis on penalty points, actions to choose which specific card to play were selected, with feedback determined by the penalty points.The learning rate for updating weights was changed so that its value was determined based on the penalty points.Separate strategies were adopted for the lead player and the other players, and multiple FALCONs were employed to improve adaptation to the game situation.Experiments demonstrated that our approach is superior to a rule-based approach.