This repository provides an OpenAI Gym interface to StarCraft: BroodWars online multiplayer game. This package was used in experiments for ICLR 2019 paper for IC3Net: Learning when to communicate at scale in multiagent cooperative and competitive tasks

• Includes a base environment which can be easily extended to various specific environment and doesn't assume anything about the downstream task.
• Both enemy and our unit's commands can be overriden for scenarios in which either of units need to be controlled for deterministic settings.
• Includes an example derived class for M vs N environment which can be further extended for specific cases of general M vs N scenarios.
• In M vs N environment, any unit type with any quantity can be initialized anywhere on map or within a specific bounding box. This environment can also be used in case of StarCraft community-building tasks as buildings themselves are units.
• Includes an explore mode environment, where one can test exploration mode for agents in big and dynamic map of StarCraft.
• Attack closest and random agent included as an example agent implementation to be used with environment.
• MvN example supports partial observable setting in which vision is limited as in fog of war.
• Supports built-in, attack-closest and attack-weakest AI strategies.

• First, we need to install TorchCraft with OpenBW support. We have written down a wiki for this.
• Make sure to install python bindings for TorchCraft as well.
• Run python setup.py develop for installing this repo.

• First, update the config present in gym-starcraft/envs/config.yml as per your requirements. Explanation is present in comments.
• Make sure all prerequisites are completed.
• To run a sample attack_closest match between 10 marines and 3 zealots in bounding box of (100, 100) to (150, 150) with GUI, run the following command:

 python examples/attack_closest.py --server_ip 127.0.0.1 --torchcraft_dir=TORCHCRAFT_DIR --set_gui --nagents 10 --max_steps 200 --frame_skip 8 --nenemies 3 --our_unit_type 0 --enemy_unit_type 65 --init_range_start 100 --init_range_end 150 --full_vision --unlimited_attack_range --initialize_enemy_together --step_size 16

Most of the other flags are self explanatory.

Use python examples/attack_closest.py -h for other options that are available.

• First, decide whether you can use either of combat MvN or explore mode environment as a start point to develop your custom environment. If you can do that, derive your new environment by extending one of these classes otherwise extend StarCraftBaseEnv like below:

import gym_starcraft.envs.starcraft_base_env as sc

class YourCustomSCEnv(sc.StarCraftBaseEnv):
    def __init__(self, args):
        # Either use argparse namespace and pass as dict
        # or pass each of the arguments using specific keywords
        super(YourCustomSCEnv, self).__init__(**vars(args))

• Now you will need to update or implement some of the required functions. In case you are using either of MvN or explore mode, then you may skip some of these so they default to original implementation. Otherwise, you need to implement most of the required functions which we list below one by one. For each of the function, see sample implementation in MvN environment:

  • First, implement _set_units function in which you set self.my_unit_pairs and self.enemy_unit_pairs which are used to instantiate our and enemy units.
  • Second, implement _action_space and _observation_space if required.
  • Third, implement _make_commands function which takes actions as parameter and returns a list of commands in TorchCraft format. See sample implementation for an example.
  • Now, implement _make_observation function which return an numpy array of shape defined in _observation_space function.
  • _has_step_completed function is checked to make sure current step is completed in _step function by default. This can be implemented in case you need to make custom checks.
  • _compute_reward function must return reward for current step in case you are planning to use it. See attack_closest agent to see how reward is retrieved for each agent from environment.
  • reward_terminal function is used to calculate reward at the end of the episode and can be called by the trainer.
  • step function implemented as per gym specification must call internal _step at some point to calculate observation from StarCraft.
  • reset function implemented as per gym specification must call internal _reset at some point to reset the actual StarCraft environment through BWAPI and return initial observation.
  • Other functions include _get_info which returns info for current step and _get_enemy_commands which can be overriden to implement custom AI for StarCraft.

Initial implementation of this package was based out of Alibaba's gym-starcraft which didn't work properly with latest TorchCraft version. To handle customized needs for our ICLR 2019 paper (not public yet), we developed this version. Without contributions of @ebetica (Zeming Lin), @tesatory (Sainbayar Sukhbaatar) and @tshrjn (Tushar Jain) and others at Facebook AI Research (FAIR) this package won't be possible.

If you use this project in your research, kindly cite it using:

@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}
}

• IC3Net code is available at IC3Net/IC3Net
• ic3net-envs is available at apsdehal/ic3net-envs

Code for this project is available under MIT license.

• Support for heterogenuous agents requires some changes in vision range calculation (Vision range is calculated only once)