Hindsight-Experience-Replay

This repository provides the Pytorch implementation of Hindsight Experience Replay on Deep Q Network and Deep Deterministic Policy Gradient algorithms.

Link to the paper: https://arxiv.org/pdf/1707.01495.pdf

Authors: Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, Pieter Abbeel, Wojciech Zaremba

Training

• You can train the model simply by running the main.py files.

  DQN With HER -> HERmain.py

  DDPG With HER -> DDPG_HER_main.py

  DQN Without HER -> main.py

• You can set the hyper-parameters such as learning_rate, discount factor (gamma), epsilon, and others while initializing the agent variable in the above-mentioned files

Running the pre-trained model

• Just run the files mentioned in the Training section with making the load_checkpoint variable to True which will load the saved parameters of the model and output the results. Just update the paths as per the saved results path.

Results

With average	Without average (contains spikes)

References

• Continuous Control With Deep Reinforcement Learning paper
• Reinforcement Learning with Hindsight Experience Replay blog
• OpenAI Spinning Up
• HER implementation in tensorflow
• OpenAI baselines