This is incomplete implementation of Shodai AlphaGo (AlphaGo Fan paper). Code was written in 2017 inspired by Ray and RocAlphaGo.
CGOS rating does not reach 2600.
http://www.yss-aya.com/cgos/19x19/cross/take.html
http://www.yss-aya.com/cgos/19x19/cross/mishima-0.1.html
(Re)development branch
python3 setup.py build_ext -i
# convert SGFs
python3 -m bamboo.scripts.policy_feature -o /path/to/feature_planes.h5 -d /path/to/sgf/directory
# run training
python3 -m bamboo.scripts.keras_supervised_policy_trainer train /path/to/weights/saved /path/to/feature_planes.h5
# Response Pattern (12-point diamond)
python3 -m bamboo.scripts.rollout_pattern -o /path/to/d12_rsp.csv -p d12_rsp -d /path/to/sgf/directory
# Non-Response Pattern (3x3)
python3 -m bamboo.scripts.rollout_pattern -o /path/to/x33.csv -p x33 -d /path/to/sgf/directory
# Non-Response Pattern (12-point diamond)
python3 -m bamboo.scripts.rollout_pattern -o /path/to/d12.csv -p d12 -d /path/to/sgf/directory
# convert SGFs
python3 -m bamboo.scripts.rollout_feature -o /path/to/rollout/feature.h5 -d /path/to/sgf/directory -p rollout
# run training
python3 -m bamboo.scripts.supervised_rollout_trainer -p rollout /path/to/rollout/feature.h5 /path/to/weights/saved
# convert SGFs
python3 -m bamboo.scripts.rollout_feature -o /path/to/tree/feature.h5 -d /path/to/sgf/directory -p tree
# run training
python3 -m bamboo.scripts.supervised_rollout_trainer -p tree /path/to/tree/feature.h5 /path/to/weights/saved
Mastering the game of Go with deep neural networks and tree search
Mastering the Game of Go without Human Knowledge
Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model