ssashir06/muzero-http-service repository overview

⁠MuZero http service

⁠Purpose

This program is created to use MuZero's game steps by another program. Mainly board games are supposed to be used.

⁠Notices

• This program provides http service to see MuZero's action.
• This program uses MuZero General⁠'s code in a submodule's directory.
• This program uses a code based on MuZero's one.
  • self_play_replicate.py⁠ is like self_play.py

⁠Docker image

Built image is located in Docker Hub⁠.

1. docker run --rm -it -p 8000:8000 ssashir06/muzero-http-service
2. Open http://localhost:8000/docs⁠ to see what apis are provided.

⁠How to build and run

⁠Requirements

• Docker for Desktop

⁠Run

⁠Using Docker Compose

You can build it manually.

1. git clone https://github.com/ssashir06/muzero-http-service.git
2. cd muzero-http-service
3. git submodule init
4. git submodule update --remote
5. docker-compose up --build

⁠Play Tic-Tac-Toe to try

⁠By using sample UI application

This is the easiest way to use it for test.

1. Open http://localhost:8888⁠ to open a sample UI application.
2. Click the "Reset" button.
3. Click the "self" button to see the MuZero's action.
4. Click the "expert" or the "random" button to see another action.
5. Click a number at the "Actions:" label or a cell in the board to select human action.

⁠By using Swagger

This might be complicated a little.

1. Open http://localhost:8000/docs⁠ to see a page of Swagger.
2. Try to use /game/{game_name}
   • :point_right: This method shows basic info for the specified game. It includes action_space and players data.
   1. Fill game_name as tictactoe
3. Try to use /game/{game_name}/{seed}
   • :point_right: This method shows the initial state of the game based on specified random seed.
   1. Fill game_name as tictactoe
   2. Fill seed as 1000
4. Try to use /game/{game_name}/{seed}/action for the 1st game step
   • :point_right: This method creates a step for the game, and shows the step's result.
   1. Fill game_name as tictactoe
   2. Fill seed as 1000
   3. Fill opponent as self
   4. Copy the steps in the Response body for the 2nd game step.

[
  {
    "action": 3,
    "root": null,
    "next": 1
  }
]

1. Try to use /game/{game_name}/{seed}/action for the 2nd game step
   • :point_right: This method creates a step for the game, and shows the step's result.
   1. Fill game_name as tictactoe
   2. Fill seed as 1000
   3. Fill opponent as self
   4. Paste steps of the 1st game step to the Request body

{
  "steps": [
    {
      "action": 3,
      "root": null,
      "next": 1
    }
  ]
}

⁠Customize

This program uses tic-tac-toe's module in the models⁠ directory. Another game's model can be used by adding or changing these things by editing docker-compose.yml⁠ file.

• Environment variable
  • :point_right: MODEL_PATH_tictactoe=./models/_for_test/tictactoe/model.checkpoint in docker-compose.yml⁠
  • The checkpoint model file can be referenced by the program by this environment variable.
  • A new environment variable having the format of MODEL_PATH_{game_name}.

⁠Issues

• The attached model file in this git repository for the Tic-Tac-Toe is not trained at all. This is just only for testing.
• If the "random" is selected as the opponent, it selects a similar action each time because it is using same random seed.

⁠References

• MuZero General⁠