Go9x9 environment


#1

Hi,

I’m new here. I’m trying to write a Go 9x9 agent by using Monte Carlo Tree Search. For this, I will need to simulate the Game of Go. First, I have some specific and very basic questions regarding the environment.

When trying with the code provided in the documentation (https://gym.openai.com/docs) and changing the environment to env = gym.make(‘Go9x9-v0’), I found out that the moves are not random. Does that suggest action = env.action_space.sample() is deterministic?

I forked the code in https://gym.openai.com/evaluations/eval_4hNanao8SIGtvddOSYwU9w and specified video_callable=always_true, env_ids=[‘Go9x9-v0’] in EnvRunner in random_agent.py. When I run it, it does not generate any video (mp4) but only some stats.json, manifest.json, meta.json, video0000x.json files, different from other environments in which I see some mp4 files. How can I generate videos?

Where can I find documentation related to Go9x9 (pachi_py)? To simulate the game, I only found out that I can do b = env.state.board.clone() and b.play_inplace(coord, color). I’d like to know how to get access to some other information, such as the number of captures, the komi, the handicap, whether the game has terminated, etc.

Does anyone have any clue?


#2

Hello,

The env.action_space.sample() should return random actions. What do you mean when you say that the moves are not random?

The idea behind Gym is that you should not request extra information from the environments other than what you see in the observations and rewards. For Go, this would be what you seen on the board. However, you can see some information inside env.state.board, which is a PachiPy object representing the board state. You can see the implemented methods here: https://github.com/openai/pachi-py/blob/master/pachi_py/cypachi.pyx#L159


#3

Thanks a lot. Very helpful information.

I don’t need to sample random actions any more so I didn’t dig into it.


#4

What is the relationship with gym boardgame package and pachi-py?