@inproceedings{32,
  author = {Sanghwa Lee and Jaeyoung Lee and Ichiro Hasuo},
  title = {Predictive PER: Balancing Priority and Diversity towards Stable Deep Reinforcement Learning},
  abstract = {<p>Prioritized experience replay (PER) samples important transitions, rather than uniformly, to improve the performance of a deep reinforcement learning agent. We\&nbsp;claim that such prioritization has to be balanced with sample diversity for making\&nbsp;the DQN stabilized and preventing forgetting. Our proposed improvement over\&nbsp;PER, called Predictive PER (PPER), takes three\&nbsp;countermeasures (TDInit, TD-Clip, TDPred) to (i) eliminate priority outliers and explosions and (ii) improve\&nbsp;the sample diversity and\&nbsp;distributions, weighted by priorities, both leading to stabilizing the DQN. The most notable among the three is the introduction of the\&nbsp;second DNN called TDPred to generalize the in-distribution priorities. Ablation\&nbsp;study and full experiments with Atari games show that each countermeasure by\&nbsp;its own way and PPER contribute to successfully enhancing stability and thus\&nbsp;performance over PER.</p>
},
  year = {2020},
  journal = {Predictive PER: Balancing Priority and Diversity towards Stable Deep Reinforcement Learning},
  url = {https://sites.google.com/view/deep-rl-workshop-neurips2020/home},
}