Monte Carlo Tree Search Optimization for Go Game AI
Tina Babu, S Tejaswi, S B Anup, Srinivas, Narla Rama Krishna Reddy, S Manasvi, Dhiraj Singh · 2025
The Monte Carlo Tree Search is the name of the best search algorithm in artificial intelligence history, especially for strategic games, where its ability was proven by integration with deep learning techniques, such as Go. The abstract underlies MCTS as a decision-making algorithm that tends to balance between exploration and exploitation through random sampling and statistical analysis of game states. MCTS works in a tree-search fashion by incrementally building up this tree, in which nodes represent game states and edges represent possible actions. The algorithm runs through four phases: selection, expansion, simulation, and backpropagation. In selection, it explores the tree until it arrives at a leaf node; expansion creates new nodes; simulation executes some random moves starting from the new node; and backpropagation sends back the simulation results into the tree. This iterative process enables MCTS to update knowledge of the best promising moves based on what happens in previous iterations. This was a significant step in the application of MCTS in Go with Google DeepMind's AlphaGo, which enabled superhuman performance by combining MCTS with neural networks. This in fact allowed AlphaGo to evaluate board positions much better and to make strategic choices to surpass human abilities in the game. The successes of MCTS inspired its usage in various domains beyond games and applications, such as optimization problems and real-time decision-making scenarios.