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
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}
}
- Find
gym-starcraftat this repository: apsdehal/gym-starcraft - Find
ic3net-envsat this repository: apsdehal/ic3net-envs
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
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
- 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:
nagentsto 5max_stepsto 40visionto 1dimto 10
For hard version, change the following arguments:
nagentsto 10max_stepsto 80visionto 1dimto 20
- 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:
nagentsto 10max_stepsto 40dimto 14add_rate_minto 0.05add_rate_maxto 0.02difficultyto medium
For hard version, change the following arguments:
nagentsto 20max_stepsto 80dimto 18add_rate_minto 0.02add_rate_maxto 0.05difficultyto hard
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
Code is available under MIT license.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent