A self-learning Monte Carlo tree search algorithm for robot path planning.

Area of Science:

• Artificial Intelligence
• Robotics
• Machine Learning

Background:

• Monte Carlo Tree Search (MCTS) is a common algorithm for decision-making in complex problems.
• Traditional MCTS can be inefficient in single-player scenarios, requiring improvements in convergence speed and search efficiency.
• Existing path planning algorithms often lack continuous self-improvement capabilities.

Purpose of the Study:

• To propose a novel self-learning MCTS algorithm (SL-MCTS) for enhanced performance in single-player scenarios.
• To improve the convergence speed and search efficiency of MCTS using a neural network.
• To enable continuous improvement of problem-solving abilities through self-learning.

Main Methods:

• Developed a self-learning Monte Carlo Tree Search (SL-MCTS) algorithm.
• Integrated a two-branch neural network (PV-Network) to predict search direction and node values, replacing the MCTS rollout process.
• Implemented a self-learning mechanism that compares current model performance against historical best models to guide optimization.

Main Results:

• SL-MCTS demonstrated superior performance in robot path planning compared to traditional MCTS and single-player MCTS.
• Achieved significantly better path quality and reduced time consumption, with time usage halved compared to traditional MCTS.
• SL-MCTS performance was comparable to specialized iterative search algorithms for path planning.

Conclusions:

• SL-MCTS offers a robust and efficient approach for single-player problem-solving, particularly in path planning.
• The integration of PV-Network and self-learning significantly enhances MCTS convergence and efficiency.
• SL-MCTS provides a scalable solution for continuous improvement in complex decision-making tasks.