Add Prioritized Experience Replay for Rainbow implementation

hengyuan-hu / rainbow

A PyTorch implementation of Rainbow DQN agent

164 stars 23 forks source link

Add Prioritized Experience Replay for Rainbow implementation #1

Open Ashutosh-Adhikari opened 6 years ago

Ashutosh-Adhikari commented 6 years ago

Should we work upon adding Prioritized Experience Replay ?

hengyuan-hu commented 6 years ago

yeah you are more than welcome to implement that. I had no access to GPU recently and just get one. But I personally may not give this high priority since a2c (https://github.com/hengyuan-hu/a2c) learns much faster.

forhonourlx commented 6 years ago

Why does a2c learn much faster? Should rainbow be the fastest overall?

Ashutosh-Adhikari commented 6 years ago

Hi @hengyuan-hu , you willing to merge the request on PER ? Or do you wish we should conduct more experiments ?