search

facebookresearch / rlmeta

RLMeta is a light-weight flexible framework for Distributed Reinforcement Learning Research.
MIT License
284 stars 28 forks source link

Refactor ReplayBuffer to improve flexibility and performance #74

Closed xiaomengy closed 2 years ago

xiaomengy commented 2 years ago

This PR made the following changes.

  1. Refactor the design of ReplayBuffer as DeepMind's Reverb project to unify the design of ReplayBuffer and Prioritized Replay Buffer.
  2. Add TensorCircularBuffer to improve the performance of ReplayBuffer with fixed shape tensor data.
  3. Refactor PPO and DQN related implementations.