Optimal Policy of Multiplayer Poker via Actor-Critic Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Game Theory
• Reinforcement Learning

Background:

• Poker presents complex challenges due to imperfect information and uncertainty, mirroring real-world problems.
• Existing methods struggle with validating optimal policies in multi-player imperfect information games.

Purpose of the Study:

• To develop an effective optimal policy learning method for multi-player poker.
• To address the limitations of current approaches in handling imperfect information and uncertainty.

Main Methods:

• Utilized Actor-Critic reinforcement learning with distinct Actor and Critic networks.
• Introduced novel asynchronous policy update algorithms (APU and Dual-APU) for multi-player scenarios.

Main Results:

• The proposed methods achieved strong performance in six-player Texas hold 'em poker.
• Learned policies demonstrated steady gains compared to existing approaches.
• Training with perfect information and testing with imperfect information proved effective.

Conclusions:

• Actor-Critic reinforcement learning provides a viable approach for learning optimal policies in imperfect information games.
• The developed methods show promise for transferability to other complex, uncertain environments.
• The approach offers an explainable method for achieving approximately optimal policies.