Repository for an AI-based chess bot designed to exploit human gameplay weaknesses, set up traps, and play much more aggressively than the vast majority of the currently available chess bots. PikeBot utilizes score evaluations generated by Stockfish to make intelligent, human-like decisions, resulting in more dynamic and exciting chess games. What sets PikeBot apart from other chess bots is its commitment to utilizing human strategic weaknesses rather than pursuing optimal gameplay, leading to more engaging gameplay.
The current version of PikeBot was trained on 500 000 publicallly available games from Lichess, and showcases promising results in the task of beating human players in the fewest possible moves. When starting from a disadvantageous position, the engine has been shown to perform better than Stockfish, requiring fewer moves to win against human opponents. To see details look for the lichess_games_results section.
Google Drive Folder with Saved Models
Website for Human vs Bot Recognition
- Data -> raw PGN data, needs to be converted to .npy files
- Preprocessed_Data -> data read from PGN files
- Model_Data -> Data which can be directly fed into the model, normalized/standardized and generally cleaned data from Preprocessed_Data
Project directory structure:
PikeBot
├── alternative_approaches
│ ├── additional_experiments_2
│ └── additional_experiments_NNUE
├── archive
| └──Stockfish - Processing Tests
├── best_models
├── evaluation
| └──games
├── experiments-results
├── human_move_prediction
├── lichess_games_results
├── lichess-bot-master
├── model_loading_data
├── pikebot
├── tests
├── training_logs
├── utils
├── .gitignore
├── conda_requirements
├── config.json.default
├── environment
├── LICENSE
├── README
└── requirements
- utils/ (in any folder): folder contains code later imported into notebooks
- logs/ (in any folder): logs of training process
- alternative_approaches/: Contains different experimental approaches, including additional experiments and NNUE-based trials.
- archive/: Stores previous test results and other legacy files, including Stockfish processing tests.
- best_models/: Contains the best-performing trained models used in the project (currently moved to the google drive, look for link in the previous sections).
- evaluation/: Stores results from evaluating the model, including played games.
- experiments-results/: Records outcomes of various experimental runs performed on the model.
- human_move_prediction/: Dedicated to training models that predict human moves based on given positions.
- lichess_games_results/: Stores results of games played on Lichess, results show PikeBot outperforming Stockfish against human opponents.
- lichess-bot-master/: Contains the implementation for interfacing PikeBot with Lichess.
- model_loading_data/: Holds data needed for loading and running trained models efficiently.
- pikebot/: heuristic implementations
- tests/: Contains unit tests and integration tests for validating different components of the bot.
- training_logs/: Stores logs from training runs to analyze model performance over time.
- utils/: Contains utility scripts and helper functions used across different parts of the project.
Default dataframe columns explanation:
| Column Name | Explanation | Expected Values |
|---|---|---|
| human | if the move was made by a human or randomly, expected model output | True/False |
| player | name of the player or "bot" if the move was made by a bot | name of the player like: "cixon123" |
| elo | player's ELO score | integer value, e.g. 1397 |
| color | color of the player | "White" or "Black" |
| event | event of the game | Event as defined in Lichess, e.g. Rated Classical game |
| clock | how much time is left in the game | floating point number representing time left in a game, e.g. 64.0 |
| stockfish_score_depth_{depth} | score for a currently playing player judged using stockfish of chosen depth in {depth} | Stockfish score as int, e.g. 744 |
| stockfish_difference_depth_{move_num} | difference between last human-made position as evaluated by stockfish and the position after making a move, i.e. how advantageous/disadvantageous a move was | Stockfish score as int, e.g. 744 |
| past_move_{move_num} | past move saved as bitboard to provide context for the model | bitboard representing board state from past turns, shaped (76, 8, 8) |
| current_move | currently performed move, expected input to the model along with optional context from previous columns | bitboard representing board state after the current human/bot move, shaped (76, 8, 8) |
| current_move_str | Forsyth–Edwards Notation (FEN) representation of the current board state | String representing the board state (example shortened here) - , e.g. r2q1rk1/... |
To retrain the human move recognition model, follow these steps:
Ensure that all required dependencies are installed. You can set up the environment using Conda:
conda env create -f conda_requirements.yml
conda activate pikebot_env--- Alternatively ---
download the environment:
conda env create -n pikebot_env --file environment.yaml
conda activate pikebot_env--- Alternatively ---
Install the dependencies on Python 3.10:
pip install -r requirements.txtMove to the human_move_prediction directory. Now, you can simply run the run.py file, the results will be saved in the corresponding folders and notebooks (CUDA GPU heavily recommended):
python run.pyIf you wish to modify the training process, refer to the Jupyter notebooks in the project.
For adjusting data processing, refer to the preprocessing scripts. For model architecture modifications, check the training notebooks inside human_move_prediction/. If you want to explore or modify the dataset, refer to the Original Data section for access to raw game data.