Automatic Optimization Method for Database Indexing By Integrating Monte Carlo Tree Search and Graph Neural Network

This article delves into the research progress and challenges in the two major fields of computer Go and computer simulated air combat. In the field of Go, an innovative METPA self play learning framework is proposed to address the shortcomings of existing algorithms in training efficiency, sample utilization, and convergence speed. By introducing stage storage tree strategy, maximum entropy tree search, and Transformer architecture, the algorithm performance is significantly improved. However, the framework also faces challenges such as high storage resource consumption and increased game time. In the field of simulated aerial combat, facing the challenge of handling high-dimensional continuous action spaces, this paper proposes the PPO algorithm (PPO-MGMM) based on multivariate joint Gaussian mixture models. By simulating complex actions through combination strategies and introducing internal KL divergence regularization techniques, the algorithm’s performance and stability are significantly improved. However, PPO-MGMM also has limitations in the application of multimodal strategies and high computational complexity. Looking ahead, combining Gaussian mixture models, Monte Carlo tree search, and advanced reinforcement learning methods is expected to further expand the application of deep reinforcement learning in various fields of real life.