Learning to Play Imperfect-Information Games by Imitating an Oracle Planner

This repository contains source code for Pommerman experiments in the paper titled "Learning to Play Imperfect-Information Games by Imitating an Oracle Planner" by Rinu Boney, Alexander Ilin, Juho Kannala and Jarno Seppänen.

Dependencies:

• pommerman package from https://github.com/MultiAgentLearning/playground
• cpommerman (see cython_env for installation instructions)
• numpy 1.18.1
• pytorch 1.5.0

Oracle Planner with Full Observability

The planning results reported in the paper can be reproduced by running:

python plan.py

Optional arguments:

Parameter	Default	Description
--planner	fdts	'fdts' or 'mcts' or 'mcs'
--mab	ts	'ts' or 'ucb'
--n_simulations	100	number of planning rollouts at each time-step
--horizon	20	depth of planning rollouts in FDTS and MCS
--n_threads	5	number of games to play in parallel
--n_episodes	100	total number of games to play

Training Follower Policy with Partial Observability

The follower results reported in the paper can be reproduced by running:

python follow.py

License

This project is licensed under the terms of the GPL-3.0 License. See LICENSE file for details.