Neuroevolved Chess Engine:

An attempt to evolve a chess engine tuning chess evaluation functions evolved by Neuroevolution frameworks (CoSyNE and NEAT).

Three architectures had been evolved for 200 generations:

1. CoSyNE-based fixed topology: 1 hidden layer architecture n = {64, 128, 1}.
2. CoSyNE-based fixed topology: 2 hidden layer architectures n = {64, 64, 64, 1}.
3. NEAT-based varying network topology: 1 hidden layer architecture n = {64, 128, 1}.

The project is still under development: Neuroevolved virtual players have partially learned the optimization task and more work is needed in order to have them fully able to play properly.

Report available at: report

---

How to test:

1. Install all required libraries in a dedicated python 3.10 environment by: pip install -r requirements.txt.
2. Download and install a Stockfish executable chess engine in your local machine: (https://stockfishchess.org/).
3. Choose the PAST_GEN generation to be tested by modifying the global var.
4. Adapt the STOCKFISH_PATH global var pointing at to your local Stockfish file in test_*.py and files if you want to test the engine against Stockfish: results will be saved in /evaluation/stockfish_eval.csv if playing with _1hidd or _2hidd versions otherwise if you need to test NEAT based chess engine, run test_dhn.py: results will be stored in /evaluation/stockfish_eval_neat.csv. You can run the testing procedure by specifying the number of the generation to be tested:

python test_1hidd.py 150

Generations available: (25, 50, 75, 100, 125, 150, 175, 200).

---

Visual assessment:

• Adapt the STOCKFISH_PATH global var pointing at to your local Stockfish file in visual_test.ipynb and execute the codecell.

Currently, you can test only the 1 hidden layer architecture by specifying the PAST_GEN generation.

---

To do:

• Evolution procedure (multi-thread spawning matches to perform tournament between individual of the current population)
• Testing procedure (single-thread running due to Stockfish limitation)
• Evolving NEAT virtual players by enforcing elitims (currently disabled)
• Visual assessment of 2-hidden layers architecture and NEAT chess based engine
• Re-implementing the engine (further implementations improvements to be defined for better performance)

---

Credits:

• Disservin: python chess engine (GNU GPL-3.0 Licence) (https://github.com/Disservin/python-chess-engine/tree/master)
• Ilya Zhelyabuzhsky: Stockfish python API (MIT Licence) (https://github.com/zhelyabuzhsky/stockfish/tree/master)
• Official Stockfish (GNU GPL-3.0 Licence): (https://github.com/official-stockfish/Stockfish)

Tested with:

• Python 3.10
• Ubuntu 22.04 LTS
• Stockfish v.16