BoxesZero: An Efficient and Computationally Frugal Dots-and-Boxes Agent.

Area of Science:

• Artificial Intelligence
• Game Theory
• Computational Efficiency

Background:

• Deep reinforcement learning (DRL) has advanced game AI, exemplified by AlphaZero.
• AlphaZero's high computational demands limit its accessibility.
• Dots-and-Boxes is a strategic game with potential for AI research.

Purpose of the Study:

• To develop a computationally frugal Dots-and-Boxes agent (BoxesZero).
• To improve learning efficiency compared to existing DRL methods.
• To demonstrate high performance with limited resources.

Main Methods:

• Introduced BoxesZero, a DRL agent for Dots-and-Boxes.
• Utilized a novel "backward training" approach, starting from high-reward states.
• Integrated domain knowledge, including extended endgame theorems for Dots-and-Boxes.
• Accelerated Monte Carlo Tree Search (MCTS) using game-specific insights.

Main Results:

• BoxesZero achieved high playing strength significantly faster than AlphaZero.
• Outperformed leading open-source Dots-and-Boxes agents (PRsboxes, DabbleBoxes) with limited GPU resources.
• Attained an ELO rating comparable to AlphaZero in less training time.
• Won the 2024 Chinese Computer Game Competition in Dots-and-Boxes.

Conclusions:

• BoxesZero demonstrates a computationally efficient and effective approach to DRL in games.
• Backward training and domain knowledge integration enhance learning speed and performance.
• The agent's success validates its approach for resource-constrained AI development.