---
id: wiki-2026-0508-prioritized-experience-replay
title: Prioritized Experience Replay
category: 10_Wiki/Topics
status: duplicate
canonical_id: experience-replay
duplicate_of: "[[Experience-Replay]]"
aliases: [PER]
source_trust_level: A
confidence_score: 0.9
verification_status: redirected
tags: [duplicate, reinforcement-learning, replay-buffer]
last_reinforced: 2026-05-10
github_commit: pending
---

# Prioritized Experience Replay

> **이 문서는 [[Experience-Replay]] 의 중복본입니다.** Canonical 문서로 redirect.

## 핵심 요약 (specialization aspects)
- **Schaul et al 2016 (ICLR)**: 매 sample transitions proportional to TD-error magnitude — high-error transitions trained more often.
- **Sampling probability**: P(i) ∝ |δ_i|^α (α=0.6 typical).
- **Importance sampling weights**: w_i = (N · P(i))^(-β) — corrects bias from non-uniform sampling; β annealed 0.4 → 1.0.
- **SumTree** data structure: O(log N) sample + update.
- 매 strict superset of uniform replay; default in Rainbow DQN.

## 🔗 Graph
- 부모: [[Experience-Replay]] (canonical)
- Adjacent: [[DQN]] · [[TD-Learning]]

## 🕓 변경 이력
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | 중복 처리 — canonical [[Experience-Replay]] 로 redirect, PER specialization aspects 보존 |