Evolutionary swarm neural network game engine for Capture Go.

Area of Science:

• Artificial Intelligence
• Computational Intelligence
• Game Theory

Background:

• Evaluating board positions is crucial for computer game engines.
• Traditional brute force search and expert knowledge bases are insufficient for complex games.
• This necessitates exploring alternative AI training methods.

Purpose of the Study:

• To investigate a hybrid algorithm combining neural networks, particle swarm optimization (PSO), and evolutionary algorithms (EAs).
• To train a game board evaluator from zero knowledge using this hybrid approach.
• To demonstrate the effectiveness of the hybrid algorithm in a complex game benchmark.

Main Methods:

• A hybrid algorithm enhancing evolutionary algorithm (EA) survivors with particle swarm optimization (PSO).
• Training high-dimensional neural networks for board evaluation through self-play.
• Utilizing the benchmark game Capture Go for experimental validation.

Main Results:

• The hybrid algorithm successfully trains neural networks for effective board evaluation.
• Experimental results on Capture Go show the hybrid approach outperforms individual EA and PSO trained engines.
• The system demonstrated superior performance against a Hill-Climbing trained engine, confirming the hybrid algorithm's efficacy.

Conclusions:

• The hybrid algorithm combining neural networks, PSO, and EAs is a powerful method for training game AI.
• This approach enables effective board evaluation from zero knowledge, surpassing limitations of traditional methods.
• The demonstrated success in Capture Go suggests broad applicability in complex game AI development.