Exploring different buffer sampling techniques to improve Hindisght Experience Replay on continuous control robotic application tasks. Continous action spaces & sparse rewards.
Comparison betwween Hindsight Experience Replay and Energy-Based Hindisght Experience Prioritization, averaged across 5 random seeds each, trained on 16 CPUs. (1 epoch = 40,000 iterations)
Results improve on state-of-the-art HER implementation by Andrychowiz et.al. (2017)
Supported by the AWS Cloud Credits for Research program.
Exploring different buffer sampling techniques to improve Hindisght Experience Replay on continuous control robotic application tasks. Continous action spaces & sparse rewards.