DeepRL is a Python repository aiming at implementing deepRL agents using PyTorch. It is compatible with Tensorboard and wandb Implemented agents are: --A2C (TD and MC, n-steps is WIP) including LSTM extension --REINFORCE

You need to have swig installed for it to work.

git clone https://github.com/YannBerthelot/deeprlyb.git
cd deeprlyb
poetry install

or

pip install -i https://test.pypi.org/simple/ deeprlyb

To run the example

poetry run python src/deeprlyb.py -s tests/config.ini

otherwise you can import agents and feed them the config like:

import gym
from deeprlyb.agents import A2C

config_file = "path/to/config.ini"
config = read_config(config_file)

env = gym.make('CartPole-v1')
agent = A2C(env, config)
agent.train_MC(env, nb_timesteps)
agent.test(env, nb_episodes_test, render=True)

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

MIT

• Add normalization and standardization of states and rewards
• Add normalization and standardization of target
• Add entropy
• Add schedules for parameters (e.g. learning rate)
• Add LSTM (multi-layers included and batch support included)
• Add rollout buffer and switch to batch learning
• Package the code into a Python lib

• Correct agent testing
• Add N-steps

• Re-add n-step A2C (that works with batch)
• Add testing during training to select agent to save
• Rework action selection logging
• Finish adding tests (cf : https://andyljones.com/posts/rl-debugging.html)
• Add explained variance