This code accompanies the paper "Scalable Multi-Agent Model-Based Reinforcement Learning".
The repository contains MAMBA implementation as well as fine-tuned hyperparameters in configs/dreamer/optimal folder.
python3.7 is required
pip install wheel
pip install flatland-2.2.2/
pip install -r requirements.txt
Installing Starcraft:
https://github.com/oxwhirl/smac#installing-starcraft-ii
python3 train.py --n_workers 2 --env flatland --env_type 5_agents
Two environments are supported for env flag: flatland and starcraft.
To train agents with optimal parameters from the paper they should be copied from configs/dreamer/optimal/ folder to DreamerAgentConfig.py and DreamerLearnerConfig.py
The code for the environment can be found at https://github.com/oxwhirl/smac
The original code for the environment can be found at https://github.com/jbr-ai-labs/NeurIPS2020-Flatland-Competition-Solution
agentcontains implementation of MAMBAcontrollerscontains logic for inferencelearnerscontains logic for learning the agentmemorycontains buffer implementationmodelscontains architecture of MAMBAoptimcontains logic for optimizing loss functionsrunnerscontains logic for running multiple workersutilscontains helper functionsworkerscontains logic for interacting with environment
envcontains environment logicnetworkscontains neural network architectures
@inproceedings{10.5555/3535850.3535894,
author = {Egorov, Vladimir and Shpilman, Alexei},
title = {Scalable Multi-Agent Model-Based Reinforcement Learning},
year = {2022},
isbn = {9781450392136},
publisher = {International Foundation for Autonomous Agents and Multiagent Systems},
address = {Richland, SC},
booktitle = {Proceedings of the 21st International Conference on Autonomous Agents and Multiagent Systems},
pages = {381–390},
numpages = {10},
keywords = {communication, multi-agent reinforcement learning, model-based reinforcement learning},
location = {Virtual Event, New Zealand},
series = {AAMAS '22}
}