Reaver is a deep reinforcement learning agent designed for training on various StarCraft II based tasks. Main focus of Reaver is following in DeepMind's footsteps, pushing state-of-the-art of the field through the lens of playing a modern video game with human-like interface and limitations. This includes observing visual features similar (though not identical) to what a human player would perceive and choosing actions from similar pool of options a human player would have. See StarCraft II: A New Challenge for Reinforcement Learning article for more details.
Though development is research-driven, the philosophy behind Reaver API is akin to StarCraft II game itself - it has something to offer both for novices and experts in the field. For hobbyist programmers Reaver offers all the tools necessary to train DRL agents by modifying only a small and isolated part of the agent (e.g. hyperparameters). For veteran researchers Reaver offers simple, but performance-optimized codebase with modular architecture: agent, model, and environment are decoupled and can be swapped at will.
- numpy >= 1.13
- absl-py >= 0.2.2
- gin-config >= 0.1.1
- TensorFlow >= 1.10
- TensorFlow Probability >= 0.4
- StarCraft II >= 4.1.2 (instructions)
- PySC2 > 2.0.1 (NB! As of 25/11/18 you must install from source since PIP version is outdated)
If you would like to use Reaver with other supported environments, you must install relevant packages as well:
- gym >= 0.10.0
- atari-py >= 0.1.5
- mujoco-py >= 1.50.0
- roboschool >= 1.0(alternative)
It is highly recommended to use Reaver on Linux OS
, both due to stability and performance considerations.
To view results with full graphics you can save a replay of the agent on Linux and open it on Windows. This is how the video recording was made.
Easiest way to install Reaver is through the PIP
package manager:
pip install reaver
NB! Reaver specifies TensorFlow
only as a soft dependency and it will not be installed by default. This is to avoid
tensorflow
overwriting tensorflow-gpu
and vise-versa. You can install tensorflow
along with Reaver by specifying either
tf-cpu
or tf-gpu
flag with pip install
command:
pip install reaver[tf-gpu]
You can also install additional extras (e.g. gym
) through the helper flags:
pip install reaver[tf-gpu,gym,atari,mujoco]
If you plan to modify Reaver
codebase you can retain its module functionality by installing from source:
$ git clone https://github.com/inoryy/reaver-pysc2
$ pip install -e reaver-pysc2/
By installing with -e
flag Python
will now look for reaver
in the specified folder, rather than site-packages
storage.
The TensorFlow
that is distributed through PIP
is built to target as many architectures / devices as possible, which
means that various optimization flags are disabled by default. For example, if your CPU supports AVX2
(is newer than 5 years),
it is highly recommended to use a custom built TensorFlow instead. If building from source is not an option for you,
then this repository might be useful - it contains newest TensorFlow
releases built for newest CUDA / CuDNN versions, which often come with performance boosts even for older GPUs.
You can train a DRL agent with multiple StarCraft II environments running in parallel with just four lines of code!
import reaver as rvr
env = rvr.envs.SC2Env(map_name='MoveToBeacon')
agent = rvr.agents.A2C(env.obs_spec(), env.act_spec(), rvr.models.build_fully_conv, rvr.models.SC2MultiPolicy, n_envs=4)
agent.run(env)
Moreover, Reaver comes with highly configurable commandline tools, so this task can be reduced to a short one-liner!
python -m reaver.run --env MoveToBeacon --agent a2c --envs 4 2> stderr.log
With the line above Reaver will initialize the training procedure with a set of pre-defined hyperparameters, optimized specifically for the given environment and agent. After awhile you will start seeing logs with various useful statistics in your terminal screen.
| T 118 | Fr 51200 | Ep 212 | Up 100 | RMe 0.14 | RSd 0.49 | RMa 3.00 | RMi 0.00 | Pl 0.017 | Vl 0.008 | El 0.0225 | Gr 3.493 | Fps 433 |
| T 238 | Fr 102400 | Ep 424 | Up 200 | RMe 0.92 | RSd 0.97 | RMa 4.00 | RMi 0.00 | Pl -0.196 | Vl 0.012 | El 0.0249 | Gr 1.791 | Fps 430 |
| T 359 | Fr 153600 | Ep 640 | Up 300 | RMe 1.80 | RSd 1.30 | RMa 6.00 | RMi 0.00 | Pl -0.035 | Vl 0.041 | El 0.0253 | Gr 1.832 | Fps 427 |
...
| T 1578 | Fr 665600 | Ep 2772 | Up 1300 | RMe 24.26 | RSd 3.19 | RMa 29.00 | RMi 0.00 | Pl 0.050 | Vl 1.242 | El 0.0174 | Gr 4.814 | Fps 421 |
| T 1695 | Fr 716800 | Ep 2984 | Up 1400 | RMe 24.31 | RSd 2.55 | RMa 30.00 | RMi 16.00 | Pl 0.005 | Vl 0.202 | El 0.0178 | Gr 56.385 | Fps 422 |
| T 1812 | Fr 768000 | Ep 3200 | Up 1500 | RMe 24.97 | RSd 1.89 | RMa 31.00 | RMi 21.00 | Pl -0.075 | Vl 1.385 | El 0.0176 | Gr 17.619 | Fps 423 |
Reaver should quickly converge to about 25-26 RMe
(mean episode rewards), which matches DeepMind results for this environment.
Specific training time depends on your hardware. Logs above are produced on a laptop with Intel i5-7300HQ CPU (4 cores)
and GTX 1050 GPU, the training took around 30 minutes.
After Reaver has finished training, you can look at how it performs by appending --test
and --render
flags to the one-liner.
python -m reaver.run --env MoveToBeacon --agent a2c --test --render 2> stderr.log
A companion Google Colab notebook notebook is available to try out Reaver online.
Many modern DRL algorithms rely on being executed in multiple environments at the same time in parallel.
As Python has GIL, this feature must be implemented through multiprocessing.
Majority of open source implementations solve this task with message-based approach (e.g. Python multiprocessing.Pipe
or MPI
),
where individual processes communicate by sending data through IPC.
This is a valid and most likely only reasonable approach for large-scale distributed approaches that companies like DeepMind and openAI operate on.
However, for a typical researcher or hobbyist a much more common scenario is having access only to a single machine environment, whether it is a laptop or a node on a HPC cluster. Reaver is optimized specifically for this case by making use of shared memory in a lock-free manner. This approach nets significant performance boost of up to 3x speed-up in StarCraft II sampling rate (and up to 100x speedup in general case), being bottle-necked almost exclusively by GPU input/output pipeline.
The three core Reaver modules - envs
, models
, and agents
are almost completely detached from each other.
This ensures that extending functionality in one module is seamlessly integrated into the others.
All configuration is handled through gin-config and can be easily shared as .gin
files.
This includes all hyperparameters, environment arguments, and model definitions.
- Advantage Actor-Critic (A2C)
- Proximal Policy Optimization (PPO)
- Generalized Advantage Estimation (GAE)
- Rewards clipping
- Gradient norm clipping
- Advantage normalization
- Baseline (critic) bootstrapping
- Separate baseline network
When experimenting with novel ideas it is important to get feedback quickly, which is often not realistic with complex environments like StarCraft II.
As Reaver was built with modular architecture, its agent implementations are not actually tied to StarCraft II at all.
You can make drop-in replacements for many popular game environments (e.g. openAI gym
) and verify implementations work with those first:
python -m reaver.run --env CartPole-v0 --agent a2c 2> stderr.log
import reaver as rvr
env = rvr.envs.GymEnv('CartPole-v0')
agent = rvr.agents.A2C(env.obs_spec(), env.act_spec())
agent.run(env)
Currently the following environments are supported by Reaver:
- StarCraft II via PySC2 (tested on all minigames)
- openAI Gym (tested on
CartPole-v0
) - Atari (tested on
PongNoFrameskip-v0
) - Mujoco (tested on
InvertedPendulum-v2
andHalfCheetah-v2
)
Map | Reaver (A2C) | DeepMind SC2LE | DeepMind ReDRL | Human Expert |
---|---|---|---|---|
MoveToBeacon | 26.3 (1.8) [21, 31] |
26 | 27 | 28 |
CollectMineralShards | 102.8 (10.8) [81, 135] |
103 | 196 | 177 |
DefeatRoaches | 72.5 (43.5) [21, 283] |
100 | 303 | 215 |
FindAndDefeatZerglings | 22.1 (3.6) [12, 40] |
45 | 62 | 61 |
DefeatZerglingsAndBanelings | 56.8 (20.8) [21, 154] |
62 | 736 | 727 |
CollectMineralsAndGas | 2267.5 (488.8) [0, 3320] |
3,978 | 5,055 | 7,566 |
BuildMarines | -- | 3 | 123 | 133 |
Human Expert
results were gathered by DeepMind from a GrandMaster level player.DeepMind ReDRL
refers to current state-of-the-art results, described in Relational Deep Reinforcement Learning article.DeepMind SC2LE
are results published in StarCraft II: A New Challenge for Reinforcement Learning article.Reaver (A2C)
are results gathered by training thereaver.agents.A2C
agent, replicatingSC2LE
architecture as closely as possible on available hardware. Results are gathered by running the trained agent in--test
mode for100
episodes, calculating episode total rewards. Listed are the mean, standard deviation (in parentheses), and min & max (in square brackets).
Map | Samples | Episodes | Approx. Time (hr) |
---|---|---|---|
MoveToBeacon | 563,200 | 2,304 | 0.5 |
CollectMineralShards | 74,752,000 | 311,426 | 50 |
DefeatRoaches | 172,800,000 | 1,609,211 | 150 |
FindAndDefeatZerglings | 29,760,000 | 89,654 | 20 |
DefeatZerglingsAndBanelings | 10,496,000 | 273,463 | 15 |
CollectMineralsAndGas | 16,864,000 | 20,544 | 10 |
BuildMarines | - | - | - |
Samples
refer to total number ofobserve -> step -> reward
chains in one environment.Episodes
refer to total number ofStepType.LAST
flags returned by PySC2.Approx. Time
is the approximate training time on alaptop
with Inteli5-7300HQ
CPU (4 cores) andGTX 1050
GPU.
Note that I did not put much time into hyperparameter tuning, focusing mostly on verifying that the agent is capable of learning
rather than maximizing sample efficiency. For example, naive first try on MoveToBeacon
required about 4 million samples,
however after some playing around I was able to reduce it down all the way to 102,000 (~40x reduction) with PPO agent.
Mean episode rewards with std.dev filled in-between. Click to enlarge.
A video recording of the agent performing on all six minigames is available online at: https://youtu.be/gEyBzcPU5-w. In the video on the left is the agent acting in with randomly initialized weights and no training, whereas on the right he is trained to target scores.
The problem of reproducibility of research has recently become a subject of many debates in science in general, and Reinforcement Learning is not an exception. One of the goals of Reaver as a scientific project is to help facilitate reproducible research. To this end Reaver comes bundled with various tools that simplify the process:
- All experiments are saved into separate folders with automatic model checkpoints enabled by default
- All configuration is handled through gin-config Python library and saved to experiment results directory
- During training various statistics metrics are duplicated into experiment results directory
- Results directory structure simplifies sharing individual experiments with full information
To lead the way with reproducibility, Reaver is bundled with pre-trained weights and full Tensorboard summary logs for all six minigames.
Simply download an experiment archive from the releases tab and unzip onto the results/
directory.
You can use pre-trained weights by appending --experiment
flag to reaver.run
command:
python reaver.run --map <map_name> --experiment <map_name>_reaver --test 2> stderr.log
Tensorboard logs are available if you launch tensorboard --logidr=results/summaries
.
In this section you can get a birdseye overview of my high level plans for Reaver. Any help with development is of course highly appreciated, assuming contributed codebase license matches (MIT).
- Documentation
- Codebase documentation
- Extending to custom environments
- Extending to custom agents
- Setup on Google Colab
- Unit tests
- For critical features such as advantage estimation
- General basic convergence guarantees of agents / models combinations
- LSTM support
- for simpler gym environments
- for full StarCraft II environment
- Asynchronous multiprocessing
- Additional agents
- Behavior Cloning / Imitation Learning
- IMPALA + PopArt
- Ape-X
- ACKTR
- Quality of Life improvements
- Plotting utility for generating research article friendly plots
- Running & comparing experiments across many random seeds
- Copying previously executed experiment
- StarCraft II raw API support
- Support for more environments
- Multi-Agent setup support
- as a proof of concept on
Pong
environment through Gym Retro - for StarCraft II through raw API
- for StarCraft II through featured layer API
- as a proof of concept on
Reaver is a very special and subjectively cute Protoss unit in the StarCraft game universe. In the StarCraft: Brood War version of the game, Reaver was notorious for being slow, clumsy, and often borderline useless if left on its own due to buggy in-game AI. However, in the hands of dedicated players that invested time into mastery of the unit, Reaver became one of the most powerful assets in the game, often playing a key role in tournament winning games.
A predecessor to Reaver, named simply pysc2-rl-agent
, was developed as the practical part of
bachelor's thesis at the University of Tartu under the
supervision of Ilya Kuzovkin and Tambet Matiisen.
You can still access it on the v1.0 branch.
If you encounter a codebase related problem then please open a ticket on GitHub and describe it in as much detail as possible. If you have more general questions or simply seeking advice feel free to send me an email.
I am also a proud member of an active and friendly SC2AI online community, we mostly use Discord for communication. People of all backgrounds and levels of expertise are welcome to join!
If you have found Reaver useful in your research, please consider citing it with the following bibtex:
@misc{reaver,
author = {Ring, Roman},
title = {Reaver: StarCraft II Deep Reinforcement Learning Agent},
year = {2018},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/inoryy/reaver}},
}