lx10077 / rlpy

A pytorch-version implementation of RL algorithms. Now it collects TRPO, ClipPPO, A2C, GAIL and ADCV.

PyTorch implementation of reinforcement learning algorithms

Important notes

• To run mujoco environments, first install mujoco-py and suggested modified version of gym which supports mujoco 1.50.
• Make sure the version of Pytorch is at least 0.4.0.
• If you have a GPU, you are recommended to set the OMP_NUM_THREADS to 1, since PyTorch will create additional threads when performing computations which can damage the performance of multiprocessing. (This problem is most serious with Linux, where multiprocessing can be even slower than a single thread):

export OMP_NUM_THREADS=1

• Code structure: Agent collects samples; Trainer facilitates learning and training; Evaluator tests trained models in new environments. All examples are placed under config file.
• After training several agents on one environment, you can plot the training process in one figure by

python utils/plot.py --env-name <ENVIRONMENT_NAME> --algo <ALGORITHM1,...,ALGORITHMn>  --x_len <ITERATION_NUM> --save_data

Policy Gradient Methods

• Trust Region Policy Optimization (TRPO) -> config/pg/trpo_gym.py
• Proximal Policy Optimization (PPO) -> config/pg/ppo_gym.py
• Synchronous A3C (A2C) -> config/pg/a2c_gym.py

Example

python config/pg/ppo_gym.py --env-name Hopper-v2 --max-iter-num 1000 --gpu

Reference

• Khrylx/PyTorch-RL
• ikostrikov/pytorch-trpo
• openai/baselines

Results

We test the policy gradient codes in these Mujoco environments with default parameters.

Generative Adversarial Imitation Learning

• GAIL -> config/gail/gail_gym.py

To save trajectory

If you want to do GAIL but without existing expert trajectories, TrajGiver will help us generate it. However, make sure the export policy has been generated and saved (i.e. train a TRPO or PPO agent on the same environment) such that TrajGiver would automatically first find the export directory, then load the policy network and running states, and eventually run the well-trained policy on desired environment.

To do imitation learning

python config/gail/gail_gym.py --env-name Hopper-v2 --max-iter-num 1000  --gpu

Action Dependent Control Variate

• ADCV -> config/adcv/v_gym.py

Example

python config/adcv/v_gym.py --env-name Walker2d-v2 --max-iter-num 1000 --variate mlp --opt minvar --gpu

Results