Antigravity Agent
f8b21af4be
10_Wiki/Topics 대규모 정리: - 오류 캡처/미완성 stub 문서 227개 제거 - 교차폴더 중복 43클러스터 병합 (63파일 → redirect) - 링크명 정규화: 깨진 링크 수정·redirect 직결·개념 매핑 ~2,400건 - 카테고리 MOC 6개 신규 생성 - Graph 섹션 미해결 related-keyword 링크 10,058건 제거 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
180 lines
6.9 KiB
Markdown
180 lines
6.9 KiB
Markdown
---
|
||
id: wiki-2026-0508-model-free-rl-vs-model-based-rl
|
||
title: Model-Free RL vs Model-Based RL
|
||
category: 10_Wiki/Topics
|
||
status: verified
|
||
canonical_id: self
|
||
aliases: [MFRL vs MBRL, Model-Based Reinforcement Learning, World Model]
|
||
duplicate_of: none
|
||
source_trust_level: A
|
||
confidence_score: 0.9
|
||
verification_status: applied
|
||
tags: [reinforcement-learning, machine-learning, planning, world-model]
|
||
raw_sources: []
|
||
last_reinforced: 2026-05-10
|
||
github_commit: pending
|
||
tech_stack:
|
||
language: Python
|
||
framework: PyTorch/JAX
|
||
---
|
||
|
||
# Model-Free RL vs Model-Based RL
|
||
|
||
## 매 한 줄
|
||
> **"매 environment dynamics 의 learn 하나, 의 X 하나 — sample efficiency 의 vs simplicity 의 trade"**. Model-free (Q-learning, PPO) 매 reward signal 의 만으로 policy 의 update — simple 의 brittle. Model-based (Dreamer, MuZero) 매 world model 의 learn → 매 imagined rollout 의 train. 2026 의 Dreamer V3, EfficientZero, DayDreamer 의 robotics deployment — sample efficiency 의 1-2 orders.
|
||
|
||
## 매 핵심
|
||
|
||
### 매 dichotomy
|
||
- **Model-free**: $\pi(a|s)$ 또는 $Q(s,a)$ 의 직접 learn. 매 transition $p(s'|s,a)$ 의 access 의 X.
|
||
- **Model-based**: $\hat{p}(s'|s,a)$, $\hat{r}(s,a)$ 의 learn → 매 plan / imagined rollout / Dyna-style.
|
||
|
||
### 매 trade-off table
|
||
| Axis | Model-Free | Model-Based |
|
||
|---|---|---|
|
||
| Sample efficiency | Low | **High** (10-100×) |
|
||
| Compute per update | Low | **High** |
|
||
| Asymptotic perf | **Often higher** | Bounded by model error |
|
||
| Stability | **Stable** | Compounding model error |
|
||
| Transfer | Poor | **Better** (model 의 reuse) |
|
||
| Implementation | **Simple** | Complex |
|
||
|
||
### 매 modern flavors
|
||
- **Model-free**: PPO, SAC, DQN family, TD3.
|
||
- **Model-based**: Dreamer V3 (RSSM), MuZero (planning + value tree), TD-MPC2, PILCO (Gaussian process).
|
||
- **Hybrid**: MBPO (model-generated rollouts → SAC), Dyna-Q.
|
||
|
||
### 매 응용
|
||
1. Robotics (sample-efficient sim-to-real).
|
||
2. Atari/board game (MuZero).
|
||
3. Drug design (sample-efficient exploration).
|
||
4. Game NPC behavior (PPO 의 still default).
|
||
|
||
## 💻 패턴
|
||
|
||
### PPO — model-free policy gradient (gymnasium)
|
||
```python
|
||
import torch, torch.nn as nn, torch.nn.functional as F
|
||
from torch.distributions import Categorical
|
||
|
||
class ActorCritic(nn.Module):
|
||
def __init__(self, obs_dim, n_act):
|
||
super().__init__()
|
||
self.shared = nn.Sequential(nn.Linear(obs_dim, 64), nn.Tanh(),
|
||
nn.Linear(64, 64), nn.Tanh())
|
||
self.pi = nn.Linear(64, n_act)
|
||
self.v = nn.Linear(64, 1)
|
||
def forward(self, x):
|
||
h = self.shared(x)
|
||
return Categorical(logits=self.pi(h)), self.v(h).squeeze(-1)
|
||
|
||
def ppo_step(ac, opt, batch, clip=0.2, vf_c=0.5, ent_c=0.01):
|
||
dist, v = ac(batch.obs)
|
||
logp = dist.log_prob(batch.act)
|
||
ratio = torch.exp(logp - batch.logp_old)
|
||
surr1 = ratio * batch.adv
|
||
surr2 = torch.clamp(ratio, 1-clip, 1+clip) * batch.adv
|
||
pi_loss = -torch.min(surr1, surr2).mean()
|
||
v_loss = F.mse_loss(v, batch.ret)
|
||
ent = dist.entropy().mean()
|
||
loss = pi_loss + vf_c * v_loss - ent_c * ent
|
||
opt.zero_grad(); loss.backward(); opt.step()
|
||
return loss.item()
|
||
```
|
||
|
||
### Dreamer-style world model (RSSM skeleton)
|
||
```python
|
||
class RSSM(nn.Module):
|
||
def __init__(self, obs_dim, act_dim, h=200, z=32):
|
||
super().__init__()
|
||
self.gru = nn.GRUCell(z + act_dim, h)
|
||
self.prior = nn.Linear(h, 2 * z) # μ, σ
|
||
self.post = nn.Linear(h + obs_dim, 2 * z)
|
||
self.dec_obs = nn.Linear(h + z, obs_dim)
|
||
self.dec_rew = nn.Linear(h + z, 1)
|
||
|
||
def step(self, h, z, a, obs=None):
|
||
h = self.gru(torch.cat([z, a], -1), h)
|
||
pri_mu, pri_log = self.prior(h).chunk(2, -1)
|
||
if obs is not None:
|
||
po_mu, po_log = self.post(torch.cat([h, obs], -1)).chunk(2, -1)
|
||
z = po_mu + torch.exp(po_log) * torch.randn_like(po_mu)
|
||
else:
|
||
z = pri_mu + torch.exp(pri_log) * torch.randn_like(pri_mu)
|
||
return h, z, (pri_mu, pri_log)
|
||
|
||
def imagine(self, h, z, policy, T=15):
|
||
states = []
|
||
for _ in range(T):
|
||
a = policy(torch.cat([h, z], -1))
|
||
h, z, _ = self.step(h, z, a)
|
||
states.append((h, z))
|
||
return states
|
||
```
|
||
|
||
### Dyna-Q (hybrid — tabular)
|
||
```python
|
||
def dyna_q(env, n_planning=10, episodes=500, alpha=0.1, gamma=0.99, eps=0.1):
|
||
Q = defaultdict(lambda: np.zeros(env.action_space.n))
|
||
model = {} # (s,a) → (r, s')
|
||
for _ in range(episodes):
|
||
s, _ = env.reset()
|
||
done = False
|
||
while not done:
|
||
a = np.random.randint(env.action_space.n) if np.random.random() < eps \
|
||
else int(np.argmax(Q[s]))
|
||
s2, r, done, *_ = env.step(a)
|
||
Q[s][a] += alpha * (r + gamma * Q[s2].max() - Q[s][a])
|
||
model[(s, a)] = (r, s2)
|
||
for _ in range(n_planning): # 매 imagined step
|
||
(sp, ap), (rp, sp2) = random.choice(list(model.items())), None
|
||
rp, sp2 = model[(sp, ap)]
|
||
Q[sp][ap] += alpha * (rp + gamma * Q[sp2].max() - Q[sp][ap])
|
||
s = s2
|
||
```
|
||
|
||
### MuZero (planning sketch — value/policy net + MCTS)
|
||
```python
|
||
# 매 environment 의 black-box; learned (representation, dynamics, prediction) heads
|
||
# search 매 imagined trajectory 의 over MCTS — replay 매 (search policy, search value, n-step return)
|
||
# 의 train. (full impl 매 muzero_general repo)
|
||
```
|
||
|
||
## 매 결정 기준
|
||
| 상황 | Approach |
|
||
|---|---|
|
||
| Lots of cheap simulation | **Model-free** (PPO/SAC) — simpler |
|
||
| Real-robot, expensive samples | **Model-based** (Dreamer V3, TD-MPC2) |
|
||
| Discrete board game | **MuZero** — planning 의 wins |
|
||
| Continuous control benchmark | SAC or DreamerV3 |
|
||
| Fast prototype | PPO — most stable, easiest to tune |
|
||
| Long-horizon planning | Model-based + planning |
|
||
|
||
**기본값**: prototype 매 PPO. Sample 매 expensive — Dreamer V3 / TD-MPC2.
|
||
|
||
## 🔗 Graph
|
||
- 부모: [[Reinforcement Learning]]
|
||
- 변형: [[PPO]]
|
||
- Adjacent: [[World Model]] · [[Planning]]
|
||
|
||
## 🤖 LLM 활용
|
||
**언제**: trade-off explanation, algorithm choice, pseudocode skeleton.
|
||
**언제 X**: 매 hyperparameter — paper-specific 의 cross-check (Dreamer V3 매 sensitivity 의 paper 의 careful).
|
||
|
||
## ❌ 안티패턴
|
||
- **MBRL 의 default 의 reach**: 매 cheap-sim 환경 의 PPO 의 win 매 simpler.
|
||
- **Imagined rollout 의 too-long horizon**: 매 model error compounds — 5-15 step 의 typical.
|
||
- **MFRL 의 sparse reward 의 hope**: 매 exploration 의 add (RND, ICM) — 또는 의 model-based 의 switch.
|
||
- **MuZero 의 small problem 의 use**: 매 overkill — tabular Q 의 enough.
|
||
- **Single-seed report**: 매 RL variance huge — 5+ seeds 의 IQM (Agarwal et al. 2021).
|
||
|
||
## 🧪 검증 / 중복
|
||
- Verified (Sutton & Barto 2nd ed. 2018; Hafner _DreamerV3_ 2023; Schrittwieser _MuZero_ Nature 2020).
|
||
- 신뢰도 A.
|
||
|
||
## 🕓 Changelog
|
||
| 날짜 | 변경 |
|
||
|---|---|
|
||
| 2026-05-08 | Phase 1 |
|
||
| 2026-05-10 | Manual cleanup — MFRL/MBRL trade-off + DreamerV3/MuZero 정리 |
|