Monte Carlo Tree Search Methods for Chess

In this repo I will share the code for some implementations of Monte Carlo Tree Seach algorithms for playing chess.

The source code is stored in mctchess, where we define the Player class and implement multiple game algorithms.

Useful notebooks will be stored in the /notebooks folder:

• The notebook MCTS_Code.ipynb stores some useful code to implement the algorithms (the main part of it already migrated to the mctchess folder).

• The notebook UTC_comparison.ipynb generates the graphics for the UTC experiments.

Execution

In order to execute a specific algorithm, one needs to install the package:

pip install -e .

and running the desired script as:

python mctchess/minimax_implementation.py

Another alternative execution procedure consists on installing only the requirements (not the entire package) and running the desired script as follows:

python -m mctchess.minimax_implementation

from the parent folder (example for the MiniMax implementation). (Note that this way to execute the scripts may lead to import problems).

The different players are stored in the directory mctchess.players. We can use them to play a Game as follows:

from chess import Board

from mctchess.game.game import Game
from mctchess.players.monte_carlo_player import MCPlayer

p_white = MiniMaxPlayer(depth=2, add_mobility=False, ab_pruning=True)
p_black = MCPlayer(n_simulations=1000, no_pools=3)

initial_board = Board()
game = Game(p_white, p_black, board, verbose=False) # Creating the Game object
game.play_game() # game rollout
winner = board.outcome().winner # in this case it was False, corresponding to the black player (Monte-Carlo based).

In the notebooks directory there are different examples of how to use players and compare their performance.

Under development

Some interesting things to do in order to upgrade this repo and make more readable are:

• Put the evaluation arguments inside a method that calls eval function.

• Investigate the recursion problems found in RAVE.

• Implement Lichess API with different methods for online interactive playing.

• Add hashing of positions with evaluation in Game to enhance performance.

• Refactor evaluation functions and pass them as argument to MiniMax players.

• Add square value to piece evaluation (use board.piece_map() and values stored in a dict)

• Add parallel computation for minimax and for game simulation.

• Put minimax and minimax_pruned out of MiniMaxPlayer.

• Perf comparison of parallel - no parallel MC implementation (profiling in notebook?).

• Parallelize Monte-Carlo evaluation.

• Test MiniMax implementation.

• Add alpha-beta pruning.