On Experience Replay

• Main Author: Tan, Renzo Roel
• Format: text
• Published: Archīum Ateneo 2019
• Subjects: n/a
• Online Access: https://archium.ateneo.edu/theses-dissertations/411
• Institution: Ateneo De Manila University

Reinforcement learning has proved to be of great utility in myriad set- tings; execution, however, may be costly due to sampling inefficiency. Experi- ence replay is used in reinforcement learning for efficient learning by recalling past experiences. While relative merits are unclear, several experience replay algorithms, namely, combined experience replay, hindsight experience replay, and prioritized experience replay, have been crafted. In this study, one surveys the existing methods and proposes hybrid replay algorithms – with hindsight and combined experience replay and with hindsight and prioritized experience replay. A comparison of the variations of experience replay incorporated into a reinforcement learning algorithm is also proffered towards an attempt to create a novel replay technique based on prioritization variants. To close, the case of multi-agent learning is considered.