An environment for the board game Go. It is implemented using OpenAI's Gym API. It is also optimized to be as efficient as possible in order to efficiently train ML models.
# In the root directory
pip install -e .
import gym
go_env = gym.make('gym_go:go-v0', size=7, komi=0, reward_method='real')
first_action = (2,5)
second_action = (5,2)
state, reward, done, info = go_env.step(first_action)
go_env.render('terminal')
0 1 2 3 4 5 6
0 ╔═╤═╤═╤═╤═╤═╗
1 ╟─┼─┼─┼─┼─┼─╢
2 ╟─┼─┼─┼─┼─○─╢
3 ╟─┼─┼─┼─┼─┼─╢
4 ╟─┼─┼─┼─┼─┼─╢
5 ╟─┼─┼─┼─┼─┼─╢
6 ╚═╧═╧═╧═╧═╧═╝
Turn: WHITE, Game State (ONGOING|PASSED|END): ONGOING
Black Area: 49, White Area: 0
state, reward, done, info = go_env.step(second_action)
go_env.render('terminal')
0 1 2 3 4 5 6
0 ╔═╤═╤═╤═╤═╤═╗
1 ╟─┼─┼─┼─┼─┼─╢
2 ╟─┼─┼─┼─┼─○─╢
3 ╟─┼─┼─┼─┼─┼─╢
4 ╟─┼─┼─┼─┼─┼─╢
5 ╟─┼─●─┼─┼─┼─╢
6 ╚═╧═╧═╧═╧═╧═╝
Turn: BLACK, Game State (ONGOING|PASSED|END): ONGOING
Black Area: 1, White Area: 1
# In the root directory.
# Defaults to a uniform random AI opponent.
python3 demo.py
GoEnv defines the Gym environment for Go. It contains the highest level API for basic Go usage.
GoGame is the set of low-level functions that defines all the game logic of Go.
GoEnv
's high level API is built on GoGame
.
These sets of functions are intended for a more detailed and finetuned
usage of Go.
We use Trump Taylor scoring, a simple area scoring, to determine the winner. A player's area is defined as the number of empty points a player's pieces surround plus the number of player's pieces on the board. The winner is the player with the larger area (a game is tied if both players have an equal amount of area on the board).
There is also support for komi
, a bias score constant to balance the advantage of black going first.
By default komi
is set to 0.
A game ends when both players pass consecutively
Reward methods are in black's perspective
-1
- White won0
- Game is tied1
- Black won0
- Otherwiseblack area - white area
.
If black won, the reward is BOARD_SIZE**2
.
If white won, the reward is -BOARD_SIZE**2
.
If tied, the reward is 0
.The state
object that is returned by the reset
and step
functions of the environment is a
6 x BOARD_SIZE x BOARD_SIZE
numpy array. All values in the array are either 0
or 1
The step
function takes in the action to execute and can be in the following forms:
None
for passing