Quick to Start, Easy to Win
- Prepare your own environment
- Let’s start large-scale distributed reinforcement learning
- Get your strong AI agent!
HandyRL is a handy and simple framework based on Python and PyTorch for distributed reinforcement learning that is applicable to your own environments. HandyRL focuses on a practicable algorithm and implementation to create a strong and winning AI in competitive games. For large scale training, HandyRL provides a controllable high parallelism power according to your environment.
HandyRL is updated at the beginning of every month except for important updates. We appreciate all contributions. Please let us know if you find a bug or have a suggestion by creating an issue and a PR.
HandyRL mainly provides a policy gradient algorithm with off-policy correction.
From the perspective of stability and performance, the off-policy version policy gradient works fine in practice. So it’s a good first choice to create a baseline AI model.
You can use some off-policy variants of update methods (targets of policy and value) from traditional ones (monte carlo, TD(λ)) to novel ones (V-Trace, UPGO).
These items can be changed in config.yaml.
As a training architecture, HandyRL adopts a learner-worker style architecture like IMPALA. The learner is a brain of training which updates a model and controls the workers. The workers have two roles. They asynchronously generate episodes (trajectories) and evaluate trained models. In episode generation, self-play is conducted as default.
HandyRL supports Python3.7+. At first, copy or fork HandyRL repository to your environment. If you want to use this script in your private project, just copy the files to your project directory and modify it there.
git clone https://github.com/DeNA/HandyRL.git
cd HandyRL
Then, install additional libraries (e.g. numpy, pytorch). Or run it in a virtual environment or container (e.g. Docker).
pip3 install -r requirements.txt
To use games of kaggle environments (e.g. Hungry Geese) you can install also additional dependencies.
pip3 install -r handyrl/envs/kaggle/requirements.txt
This section shows the training a model for Tic-Tac-Toe. Tic-Tac-Toe is a very simple game. You can play by googling "Tic-Tac-Toe".
Set config.yaml for your training configuration. When you run a training with Tic-Tac-Toe and batch size 64, set like the following:
env_args:
env: 'TicTacToe'
train_args:
...
batch_size: 64
...NOTE: Here is the list of games implemented in HandyRL. All parameters are shown in Config Parameters.
After creating the configuration, you can start training by running the following command. The trained models are saved in models folder every update_episodes described in config.yaml.
python main.py --train
After training, you can evaluate the model against any models. The below code evaluate the model of epoch 1 for 100 games with 4 processes.
python main.py --eval models/1.pth 100 4
NOTE: Default opponent AI is random agent implemented in evaluation.py. You can change the agent with any of your agents.
- Config Parameters shows a list of parameters of
config.yaml. - Large Scale Training is a procedure for large scale training remotely.
- Train with Customized Environment explains an interface of environment to create your own game.
- API shows entry-point APIs of
main.py
