Enhancing the AlphaGo Zero Algorithm Through Kolmogorov-Arnold Networks
Jianuo Lei, Hang Ouyang, Yang Tan, Qing Li, Haolan Wang, Kaixin Deng · 2024
With the rapid advancement of artificial intelligence, AlphaGo has revolutionized the game of Go, ultimately evolving into AlphaGo Zero. Despite its success, the neural network architecture of the AlphaGo Zero model, based on ResNet, is characterized by a high number of parameters and computational complexity, making it challenging to deploy on low-power devices. To address this issue, this paper presents an adaptation of the AlphaGo Zero algorithm by substituting its ResNet with a Kolmogorov-Arnold Network (KAN) structure. This modification is applied and evaluated within the game "Dots-and-boxes", serving as a test case for potential real-world applications. The paper details the design and training process of the KAN-based model, alongside a rigorous comparison with the original AlphaGo Zero model. Experimental results indicate that the KAN-based model not only preserves competitive performance in Dots-and-boxes but also achieves a substantial reduction in model complexity and parameter count. These improvements demonstrate the model's enhanced suitability for resource-constrained environments, making it more efficient for low-power device deployment.