Reward Prediction Error Prioritised Experience Replay (RPE-PER)

The following repository contains the PyTorch implementation of Reward Prediction Error Prioritisation Experience Replay (RPE-PER). It is integrated into two off-policy RL algorithms: TD3 and SAC.

The algorithm is tested on MuJoCo continuous control suite.

Network Architecture

Instructions

Prerequisite Versions

Library	Version
pydantic	1.10.10
MuJoCo	2.3.3

Training

To train the RD-PER TD3 or SAC, use the following command:

python3 training_loop_SAC.py
# or
python3 training_loop_TD3.py

Citation

Please cite the paper and the github repository if used.