RLlib Multi Agent Self-Play Example - MultiPlayer Gridman

In this version of the GridMan environment you train gridman but also his enemies!

GridMan and the 3 enemies all have completely seperate neural networks and separate rewards.

More documentation can be found here

This repository can be used as a starter for training any of the Griddly environemnts using RLLib

Installing Dependencies

We use poetry to manage the dependencies of this project. You can set up a poetry environment with the command:

poetry install

You can then activate the poetry environment using:

poetry shell

Training the example environment

python train.py

Options for training

There are three variables that you can change in train.py:

environment_name = "TestEnvironment"
environment_yaml = "gridman/gridman_multiagent.yaml"
gridman_model_name = "SimpleConvAgent"
enemy_model_name = "SimpleConvAgent"

environment_name

The name of the environment

environment_yaml

the yaml file containing the GDY of the environment

model_name

We provide two simple models that can be used are SimpleConvAgent and GlobalAveragePoolingAgent

Weights and Biases

You can find training information here:

https://wandb.ai/griddlyai/RLLib%20Gridman%20MultiAgent

Citing

If you use this environment, please cite the original Griddly Paper:

@article{
  author    = {Chris Bamford and
               Shengyi Huang and
               Simon M. Lucas},
  title     = {Griddly: {A} platform for {AI} research in games},
  journal   = {CoRR},
  volume    = {abs/2011.06363},
  year      = {2020},
  url       = {https://arxiv.org/abs/2011.06363},
  eprinttype = {arXiv},
  eprint    = {2011.06363},
}