a0-jax

AlphaZero in JAX using deepmind mctx library.

pip install -r requirements.txt

Train agent

Connect-Two game

python train_agent.py --weight-decay=1e-2 --num-iterations=3

Connect-Four game

TF_CPP_MIN_LOG_LEVEL=2 \
python train_agent.py \
    --game_class="games.connect_four_game.Connect4Game" \
    --agent_class="policies.resnet_policy.ResnetPolicyValueNet" \
    --batch-size=4096 \
    --num_simulations_per_move=32 \
    --num_self_plays_per_iteration=102400 \
    --learning-rate=1e-2 \
    --num_iterations=500 \
    --lr-decay-steps=200000

A live Connect-4 agent is running at https://huggingface.co/spaces/ntt123/Connect-4-Game. We use tensorflow.js to run the policy on the browser.

Caro (Gomoku) game

TF_CPP_MIN_LOG_LEVEL=2 \
python3 train_agent.py \
    --game-class="games.caro_game.CaroGame" \
    --agent-class="policies.resnet_policy.ResnetPolicyValueNet128" \
    --selfplay-batch-size=1024 \
    --training-batch-size=1024 \
    --num-simulations-per-move=32 \
    --num-self-plays-per-iteration=102400 \
    --learning-rate=1e-2 \
    --random-seed=42 \
    --ckpt-filename="./caro_agent_9x9_128.ckpt" \
    --num-iterations=100 \
    --lr-decay-steps=500000

A live Caro agent is running at https://caro.ntt123.repl.co.

Go game

TF_CPP_MIN_LOG_LEVEL=2 \
python3 train_agent.py \
    --game-class="games.go_game.GoBoard9x9" \
    --agent-class="policies.resnet_policy.ResnetPolicyValueNet128" \
    --selfplay-batch-size=1024 \
    --training-batch-size=1024 \
    --num-simulations-per-move=32 \
    --num-self-plays-per-iteration=102400 \
    --learning-rate=1e-2 \
    --random-seed=42 \
    --ckpt-filename="./go_agent_9x9_128.ckpt" \
    --num-iterations=200 \
    --lr-decay-steps=1000000

A live Go agent is running at https://go.ntt123.repl.co. You can run the agent on your local machine with the go_web_app.py script.

We also have an interative colab notebook that runs the agent on GPU to reduce inference time.

Plot the search tree

python plot_search_tree.py 
# ./search_tree.png

Play

python play.py

TPU sponsor

Agents in the above demos are trained on Google TPUs sponsored by Google under the TPU Research Cloud program.