PyTorch implementation of Distributed Distributional Deterministic Policy Gradients (https://arxiv.org/abs/1804.08617).

D4PG and D3PG implementations with following features

• learner, sampler and agents run in separate processes
• exploiter agent(s) exists which acts without noise in actions on target network
• GPU is hold only by exploiters, all other exploration processes are run on CPU

Project was tested on Ubuntu 18.04, Intel i5 with 4 cores, Nvidia GTX 1080Ti

Run python train.py --config configs/openai/d4pg/walker2d_d4pg.yml

python -m unittest discover

Configs for reproducing curves below can be found in configs directory (num parallel agents = 4).

OpenAI Mujoco

DMControl

All results were obtained with configs in configs directory

• Continuous control with deep reinforcement learning, [https://arxiv.org/abs/1509.02971]
• Distributed Distributional Deterministic Policy Gradients [https://arxiv.org/abs/1804.08617]