EasyDeepLearn

How do you size a replay buffer?

medium

Answer

  • Trade-off: bigger buffer → more diverse data, older stale off-policy data.
  • Rule: enough to cover diverse experience but not so large that stale data hurts.
  • Atari DQN: 1M transitions.
  • Continuous control: 10610710^{6} - 10^{7}.
  • Off-policy sensitivity: SAC works with larger buffers; DDPG less so.
  • Prioritized replay adjusts distribution automatically.
  • Off-policy correction (V-trace, ReTrace) helps with staleness in on-policy-ish algorithms.
Check yourself — multiple choice
  • Random
  • Bigger buffer = more diversity but more staleness; DQN-Atari 1M, continuous control 10610710^{6} - 10^{7}; prioritized replay + off-policy correction (V-trace) mitigate staleness
  • Not needed
  • Only tiny

Replay size: diversity vs staleness; Atari 1M; continuous control up to 10710^{7}.

#value-methods#engineering

Practise Reinforcement Learning

214 interview questions in this topic.

Related questions