EasyDeepLearn

Off-policy policy gradient — how does importance sampling enter?

hard

Answer

  • Data collected under behavior πb{\pi}_{b}, but want to update target π_θ: use importance weight ρ = π_θ(as)/πb(as)(a \mid s) / {\pi}_{b}(a \mid s) → ∇J ≈ E_π_b[ρ ∇log π_θ · A].
  • Weight variance explodes when π_θ far from πb{\pi}_{b}.
  • Solutions: clip ratio (PPO), truncate importance weights (Retrace, V-trace in IMPALA), or use deterministic policy gradient (no IS needed — DDPG/SAC).
Check yourself — multiple choice
  • Same as on-policy
  • IS weight ρ = π_θ/πb{\pi}_{b} corrects for behavior/target mismatch; variance explodes if policies diverge → clip (PPO) / truncate (V-trace, Retrace) / deterministic PG (no IS)
  • Random
  • Not real

Off-policy PG: IS weight π_θ/πb{\pi}_{b}; clip / truncate / deterministic PG.

#policy-methods#theory

Practise Reinforcement Learning

214 interview questions in this topic.

Related questions