Schokoban is a Sokoban solver, which unlike state-of-the-art Sokoban solvers that often use heuristic or search-based algorithms, experiments with applying Monte Carlo Tree Search (MCTS) to this classic puzzle game.
The project explores how to adapt MCTS to handle the unique challenges of Sokoban, such as redundant states and branching complexity. A key innovation is a new method for managing redundant states within the tree search, resulting in substantial performance improvements over a baseline approach.
While the solver’s main goal is experimental research rather than setting new performance records, it can still successfully solve most puzzles in well-known level collections like Microban III.
If you want to learn more, see performance results, feel free to explore the code on GitHub: https://github.com/paulkroe/Schokoban
Example Sokoban level from by Skinner.