IC3Net

This repository contains reference implementation for IC3Net paper (accepted to ICLR 2019), Learning when to communicate at scale in multiagent cooperative and competitive tasks, available at https://arxiv.org/abs/1812.09755

Cite

If you use this code or IC3Net in your work, please cite the following:

@article{singh2018learning,
  title={Learning when to Communicate at Scale in Multiagent Cooperative and Competitive Tasks},
  author={Singh, Amanpreet and Jain, Tushar and Sukhbaatar, Sainbayar},
  journal={arXiv preprint arXiv:1812.09755},
  year={2018}
}

Standalone environment version

• Find gym-starcraft at this repository: apsdehal/gym-starcraft
• Find ic3net-envs at this repository: apsdehal/ic3net-envs

Installation

First, clone the repo and install ic3net-envs which contains implementation for Predator-Prey and Traffic-Junction

git clone https://github.com/IC3Net/IC3Net
cd IC3Net/ic3net-envs
python setup.py develop

Optional: If you want to run experiments on StarCraft, install the gym-starcraft package included in this package. Follow the instructions provided in README inside that packages.

Next, we need to install dependencies for IC3Net including PyTorch. For doing that run:

pip install -r requirements.txt

Running

Once everything is installed, we can run the using these example commands

Note: We performed our experiments on nprocesses set to 16, you can change it according to your machine, but the plots may vary.

Note: Use OMP_NUM_THREADS=1 to limit the number of threads spawned

Predator-Prey

• IC3Net on easy version

python main.py --env_name predator_prey --nagents 3 --nprocesses 16 --num_epochs 2000 --hid_size 128 --detach_gap 10 --lrate 0.001 --dim 5 --max_steps 20 --ic3net --vision 0 --recurrent

• CommNet on easy version

python main.py --env_name predator_prey --nagents 3 --nprocesses 16 --num_epochs 2000 --hid_size 128 --detach_gap 10 --lrate 0.001 --dim 5 --max_steps 20 --commnet --vision 0 --recurrent

• IC on easy version

python main.py --env_name predator_prey --nagents 3 --nprocesses 16 --num_epochs 2000 --hid_size 128 --detach_gap 10 --lrate 0.001 --dim 5 --max_steps 20 --vision 0 --recurrent

• IRIC on easy version

python main.py --env_name predator_prey --nagents 3 --nprocesses 16 --num_epochs 2000 --hid_size 128 --detach_gap 10 --lrate 0.001 --dim 5 --max_steps 20 --mean_ratio 0 --vision 0 --recurrent

For medium version, change the following arguments:

• nagents to 5
• max_steps to 40
• vision to 1
• dim to 10

For hard version, change the following arguments:

• nagents to 10
• max_steps to 80
• vision to 1
• dim to 20

Traffic Junction

• IC3Net on easy version

python main.py --env_name traffic_junction --nagents 5 --nprocesses 16 --num_epochs 2000 --hid_size 128 --detach_gap 10 --lrate 0.001 --dim 6 --max_steps 20 --ic3net --vision 0 --recurrent --add_rate_min 0.1 --add_rate_max 0.3 --curr_start 250 --curr_end 1250 --difficulty easy

• CommNet on easy version

python main.py --env_name predator_prey --nagents 5 --nprocesses 16 --num_epochs 2000 --hid_size 128 --detach_gap 10 --lrate 0.001 --dim 6 --max_steps 20 --commnet --vision 0 --recurrent  --add_rate_min 0.1 --add_rate_max 0.3 --curr_start 250 --curr_end 1250 --difficulty easy

• IC on easy version

python main.py --env_name predator_prey --nagents 5 --nprocesses 16 --num_epochs 2000 --hid_size 128 --detach_gap 10 --lrate 0.001 --dim 6 --max_steps 20 --vision 0 --recurrent  --add_rate_min 0.1 --add_rate_max 0.3 --curr_start 250 --curr_end 1250 --difficulty easy

• IRIC on easy version

python main.py --env_name predator_prey --nagents 5 --nprocesses 16 --num_epochs 2000 --hid_size 128 --detach_gap 10 --lrate 0.001 --dim 6 --max_steps 20 --mean_ratio 0 --vision 0 --recurrent --add_rate_min 0.1 --add_rate_max 0.3 --curr_start 250 --curr_end 1250 --difficulty easy

For medium version, change the following arguments:

• nagents to 10
• max_steps to 40
• dim to 14
• add_rate_min to 0.05
• add_rate_max to 0.02
• difficulty to medium

For hard version, change the following arguments:

• nagents to 20
• max_steps to 80
• dim to 18
• add_rate_min to 0.02
• add_rate_max to 0.05
• difficulty to hard

StarCraft

Make sure you have gym-starcraft properly installed and configuration properly configured.

For explore task 50x50, 10Medic, see the examples below, replace torchcraft_dir argument with your torchcraft directory location

• IC3Net

python -u main.py --env_name starcraft --task_type explore --nagents 10 --num_epochs 1000 --hid_size 128 --lrate 0.002 --max_steps 60 --nprocesses 16 --torchcraft_dir=~/Public/TorchCraft --frame_skip 8 --nenemies 1 --our_unit_type 34 --enemy_unit_type 34 --init_range_end 150 --ic3net --recurrent --rnn_type LSTM --detach_gap 10 --stay_near_enemy --explore_vision 10 --step_size 16

• CommNet

python -u main.py --env_name starcraft --task_type explore --nagents 10 --num_epochs 1000 --hid_size 128 --lrate 0.002 --max_steps 60 --nprocesses 16 --torchcraft_dir=~/Public/TorchCraft --frame_skip 8 --nenemies 1 --our_unit_type 34 --enemy_unit_type 34 --init_range_end 150 --commnet --recurrent --rnn_type LSTM --detach_gap 10 --stay_near_enemy --explore_vision 10 --step_size 16

• IRIC

python -u main.py --env_name starcraft --task_type explore --nagents 10 --num_epochs 1000 --hid_size 128 --lrate 0.002 --max_steps 60 --nprocesses 16 --torchcraft_dir=~/Public/TorchCraft --frame_skip 8 --nenemies 1 --our_unit_type 34 --enemy_unit_type 34 --init_range_end 150 --mean_ratio 0 --recurrent --rnn_type LSTM --detach_gap 10 --stay_near_enemy --explore_vision 10 --step_size 16

• IC

python -u main.py --env_name starcraft --task_type explore --nagents 10 --num_epochs 1000 --hid_size 128 --lrate 0.002 --max_steps 60 --nprocesses 16 --torchcraft_dir=~/Public/TorchCraft --frame_skip 8 --nenemies 1 --our_unit_type 34 --enemy_unit_type 34 --init_range_end 150 --recurrent --rnn_type LSTM --detach_gap 10 --stay_near_enemy --explore_vision 10 --step_size 16

For 75x75, set --init_range_end to 175.

For Combat version:

• IC3Net

python -u main.py --env_name starcraft --task_type combat --nagents 10 --num_epochs 1000 --hid_size 128 --lrate 0.002 --max_steps 60 --nprocesses 16 --torchcraft_dir=~/Public/TorchCraft --frame_skip 8 --nenemies 3 --our_unit_type 0 --enemy_unit_type 65 --init_range_end 150 --ic3net --recurrent --rnn_type LSTM --detach_gap 10 --explore_vision 10 --step_size 16

• CommNet

python -u main.py --env_name starcraft --task_type combat --nagents 10 --num_epochs 1000 --hid_size 128 --lrate 0.002 --max_steps 60 --nprocesses 16 --torchcraft_dir=~/Public/TorchCraft --frame_skip 8 --nenemies 3 --our_unit_type 0 --enemy_unit_type 65 --init_range_end 150 --commnet --recurrent --rnn_type LSTM --detach_gap 10 --explore_vision 10 --step_size 16

• IRIC

python -u main.py --env_name starcraft --task_type combat --nagents 10 --num_epochs 1000 --hid_size 128 --lrate 0.002 --max_steps 60 --nprocesses 16 --torchcraft_dir=~/Public/TorchCraft --frame_skip 8 --nenemies 3 --our_unit_type 0 --enemy_unit_type 65 --init_range_end 150 --mean_ratio 0 --recurrent --rnn_type LSTM --detach_gap 10 --explore_vision 10 --step_size 16

• IC

python -u main.py --env_name starcraft --task_type combat --nagents 10 --num_epochs 1000 --hid_size 128 --lrate 0.002 --max_steps 60 --nprocesses 16 --torchcraft_dir=~/Public/TorchCraft --frame_skip 8 --nenemies 3 --our_unit_type 0 --enemy_unit_type 65 --init_range_end 150 --recurrent --rnn_type LSTM --detach_gap 10 --explore_vision 10 --step_size 16

Contributors

• Amanpreet Singh (@apsdehal)
• Tushar Jain (@tshrjn)
• Sainbayar Sukhbaatar (@tesatory)

License

Code is available under MIT license.