Sokoban RL Solver

This project applies Reinforcement Learning algorithms to solve Sokoban puzzles. The following algorithms have been implemented:

• Monte Carlo Policy Optimization
• Temporal Difference Learning
• REINFORCE
• Actor-Critic

To solve a puzzle using one of our algorithms, run:

python .\solve_puzzle.py <puzzle_file> <algorithm>

Algorithms: monte_carlo, td_learning, reinforce, actor_critic

The following algorithms from stable_baselines3 are also available:

• Proximal Policy Optimization
• Asynchronous Advantage Actor Critic
• Deep Q-Network

To solve a puzzle using a stable_baselines3 algorithm, run:

python .\sb3_solve_puzzle.py <puzzle_file> <algorithm>

Algorithms: PPO, A2C, DQN

Examples:

python .\solve_puzzle.py .\data\puzzles\level_1.txt monte_carlo

python .\sb3_solve_puzzle.py .\data\puzzles\level_2.txt PPO

To solve a puzzle using non-RL algorithms (BFS, A*...), cd into sokoban-solver-generator (a fork of xbandrade's repository ).