Official code for the paper "Asynchronous, Option-Based Multi-Agent Policy Gradient: A Conditional Reasoning Approach"

• Water-Filling
• Tool-Delivery
• Capture-Target

1. Install conda and create a conda environment (python 3.6.8)
   conda create env --name mahrl
2. Install necessary dependencies
   pip install -r requirements.txt
3. Validate the installation of ai2thor simulator
   • On Windows, you may want to check: allenai/ai2thor#811.
   • On Linux, you may want to check: https://github.com/allenai/ai2thor-docker. You don't necessarily use the docker, but startx() and an minimal example is helpful to run the simulator on a Linux server without a monitor.
   • Anyway, run this example to see if everything about ai2thor works well.

   import ai2thor.platform
   from pprint import pprint
   
   if __name__ == '__main__':
       controller = ai2thor.controller.Controller(platform=ai2thor.platform.CloudRendering, scene='FloorPlan28')
       event = controller.step(action='RotateRight')
       pprint(event.metadata['agent'])```
   

• Water-Filling task
  python train_wf.py --scheme [fully-dec, partial-dec, fully-cen, partial-cen, sync-cut, sync-wait, end2end] --seed [seed]
• Tool-Delivery task
  python tool_delivery/train_td.py --scheme [fully-dec, partial-dec, fully-cen, partial-cen, sync-cut,sync-wait] --seed [seed]
• Capture-Target task
  python capture_target/train_ct.py --scheme [fully-dec, fully-cen] --seed [seed]

Use the same commands above, but add one more line in train_wf.py, train_td.py or train_ct.py as follows.
For example, in train_td.py line 95, add all_args.model_dir = "./results/toolDeliverySeparate/fully-dec/mappo/mlp/run10/models", and change the model saving path to yours.

@article{lyu2022asynchronous,
title={Asynchronous, Option-Based Multi-Agent Policy Gradient: A Conditional Reasoning Approach},
author={Lyu, Xubo and Banitalebi-Dehkordi, Amin and Chen, Mo and Zhang, Yong},
journal={arXiv preprint arXiv:2203.15925},
year={2022}
}