Robotics Institute, Carnegie Mellon University
This is the official repository of the paper Natural Emergence of Heterogeneous Strategies in Artificially Intelligent Competitive Teams, presented in a spotlight talk in Robotics Science and Systems (RSS) 2020 workshop on Heterogeneous Multi-Robot Task Allocation and Coordination. Increasingly complex multiagent strategies can emerge naturally through competition in a mixed cooperative-competitive environment.
I have tested the code on Ubuntu 16.04 and 18.04. See requirements.txt
file for the list of dependencies. Create an anaconda environment or virtualenv with python 3.6 and setup everything by executing pip install -r requirements.txt
.
conda create -n fortattack python=3.6 pip
conda activate fortattack
pip install -r requirements.txt
python train_fortattack.py --save-dir tmp_3
See arguments.py
file for the list of various command line arguments one can set while running the scripts.
python train_fortattack_v2.py --train-guards-only --pretrained-guard --guard-load-dir tmp_1 --guard-ckpt 2520 --attacker-load-dir tmp_1 --attacker-ckpts 220 650 1240 1600 2520 --save-dir tmp_4 --render
python test_fortattack.py --test --load-dir tmp_1 --ckpt 220
python test_fortattack_v2.py --test --train-guards-only --num-eval-episodes 10 --load-dir tmp_2 --ckpt 5050 --attacker-load-dir tmp_1 --attacker-ckpts 220 --render
I have provided some pretrained models in marlsave
directory
-
tmp_1
: normal training results. Multiagent strategies corresponding the checkpoints are as follows. Please refer to the paper for details of these strategies.ep220
: Flash laser strategy of guardsep650
: Sneaking strategy of attackersep1240
: Spreading and flashing strategy of attackersep1600
: Sneaking strategy of attackersep2520
: Smartly spreading strategy of guards
-
tmp_2/ep5050.pt
: Guards' policy after ensemble training
If you use our code in your research, please cite our paper:
@article{deka2020natural, title={Natural Emergence of Heterogeneous Strategies in Artificially Intelligent Competitive Teams}, author={Deka, Ankur and Sycara, Katia}, journal={arXiv preprint arXiv:2007.03102}, year={2020} }
For any queries, feel free to raise an issue or contact me at adeka@cs.cmu.edu or ankurnet1996@gmail.com.
The environment is built up on MAPE
The algorithm for this repo is built up on marl_transfer
This project is licensed under the MIT License.