Multi-Objective Optimization Based on AlphaZero Method Applied to Connection Games

Open-ended learning, introduced by the DeepMind team in 2021, represents a novel AI approach to achieve multi-objective optimization, allowing it to simultaneously handle multiple tasks, whereas traditional AI is optimized only for a single task. This study describes an AI implementation similar to open-ended learning using relatively familiar technologies and game rules, making it adaptable to two different game rule sets. Experimental results indicate that, given equal training time, multi-objective AlphaZero outperforms the single-objective optimized version of AlphaZero. Training using both rule sets impacts the neural network. Our results indicate that simpler training data can accelerate the model’s initial learning phase without compromising performance across the two rule sets. Multi-objective training is found to enhance the overall efficiency and effectiveness of neural network learning. Specifically, incorporating simpler rule sets accelerates early-stage training and helps lay the foundation for learning more complex rules. The findings suggest that leveraging the interaction between different levels of rule complexity can achieve more balanced and comprehensive training outcomes, ultimately resulting in a more generalizable AI model.