sokoban

Sokoban game AI for UCI CS271 course project

Building the Environment

The Sokoban environment is implemented in C++ to ensure that the game can run quickly, and we can perform search efficiently.

In order to build the environment you will need:

• A modern C++ compiler capable of C++17 or above and OpenMP.
• Python 3.7 or above
• Cython

Run make in the root directory in order to create the environment and the python wrapper.

Python Instruction

In order to run the search algorithms, you will need the following libraries.

• Python 3.7 or above
• Numpy
• Scipy
• fastdist
• PyTorch
• CUDA (Optional but makes DQN much faster)
• TorchSpread (Optional for training networks)

Run source environment.sh in the main directory in order to be able to use all of the python features from anywhere.

Run python run.py -h in order to get a help menu for the solver code.

Heuristic Options

Activate each of these using the -s options for run.py

1. Manhattan Greedy Selection
2. Euclidean Greedy Selection
3. Manhattan Hungarian Selection
4. Euclidean Hungarian Selection
5. Small Q-Learning Network (CPU)
6. Large Q-Learning Network (GPU & Batch A* Search)