RL-Experiments

RL-Experiments aims to modify and compare deep RL algorithms in single machine easily. For distributed training, I highly recommend ray.

The codes refer to openai/baselines mostly but are implemented by PyTorch. We also highlight the differences between implementation and paper which can be found by searching highlight in codes. Evaluated on 4 atari games, our implementation is 15% faster than baseline with similar performance on average in single machine.

Dependency

• Python: 3.5+
• PyTorch: 1.0.1+

Evaluation

With the same default parameters in openai/baselines, the FPS and performance with random seed 0 on four environments are illustrated as follows.

Devices:

• 1 NVIDIA GTX 1070
• 8 Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz

Pong

SpaceInvaders

BeamRider

Seaquest

Usage

git clone https://github.com/Officium/RL-Experiments.git
cd RL-Experiments/src
python run.py --env=CartPole-v1 --env_type=classic_control --algorithm=dqn --number_timesteps=1e5

Implemented algorithms

• A2C
• DQN (with double, C51, dueling, noisy and prioritized replay)
• PPO
• REINFORCE
• TRPO