Abstract
This paper proposes a self-learning Monte Carlo tree search algorithm (SL-MCTS), which has the ability to continuously improve its problem-solving ability in single-player scenarios. SL-MCTS combines the MCTS algorithm with a two-branch neural network (PV-Network). The MCTS architecture can balance the search for exploration and exploitation. PV-Network replaces the rollout process of MCTS and predicts the promising search direction and the value of nodes, which increases the MCTS convergence speed and search efficiency. The paper proposes an effective method to assess the trajectory of the current model during the self-learning process by comparing the performance of the current model with that of its best-performing historical model. Additionally, this method can encourage SL-MCTS to generate optimal solutions during the self-learning process. We evaluate the performance of SL-MCTS on the robot path planning scenario. The experimental results show that the performance of SL-MCTS is far superior to the traditional MCTS and single-player MCTS algorithms in terms of path quality and time consumption, especially its time consumption is half less than that of the traditional MCTS algorithms. SL-MCTS also performs comparably to other iterative-based search algorithms designed specifically for path planning tasks.