youngadvance/evolutionary-reinforcement-learning

Codebase for Evolutionary Reinforcement Learning (ERL) from the paper "Evolution-Guided Policy Gradients in Reinforcement Learning" published at NeurIPS 2018

The master branch introduces many additions over the NeurIPS paper published in 2018 - improving significantly on runtime and algorithmic performance. The primary changes are detailed below:

1. Parallleized rollouts
2. Soft-Actor Critic (SAC) https://arxiv.org/abs/1801.01290 replacing the DDPG used originally
3. Support for Discrete Environments using a form of DDQN + Maximum Entropy Reinforcement Learning

Please switch to neurips_paper_2018 branch if you wish to reproduce the original results from the paper https://papers.nips.cc/paper/7395-evolution-guided-policy-gradient-in-reinforcement-learning.pdf

Dependencies Tested on

Python 3.6.9
Pytorch 1.2
Numpy 1.18.1
Gym 0.15.6
Mujoco-py v1.50.1.59

To Run

python main.py --env $ENV_NAME$

Environment name examples to get you started

Continous

'Humanoid-v2'
'Hopper-v2'
'HalfCheetah-v2'
'Swimmer-v2'
'Ant-v2'
'Walker2d-v2'
'Reacher-v2' \

Discrete

'CartPole-v1'
'Pong-ram-v0'
'Qbert-ram-v0'
'MountainCar-v0'

To use your own custom environment

Write a gym-compatible wrapper around your environment and register it with the gym runtime