Antigravity Agent
146 lines
5.1 KiB
Markdown
146 lines
5.1 KiB
Markdown
---
|
||
id: wiki-2026-0508-dopamine
|
||
title: Dopamine
|
||
category: 10_Wiki/Topics
|
||
status: verified
|
||
canonical_id: self
|
||
aliases: [Reward System, Reinforcement Signal, Prediction Error]
|
||
duplicate_of: none
|
||
source_trust_level: A
|
||
confidence_score: 0.85
|
||
verification_status: applied
|
||
tags: [neuroscience, reinforcement-learning, motivation, ux]
|
||
raw_sources: []
|
||
last_reinforced: 2026-05-10
|
||
github_commit: pending
|
||
tech_stack:
|
||
language: python
|
||
framework: rl
|
||
---
|
||
|
||
# Dopamine
|
||
|
||
## 매 한 줄
|
||
> **"매 reward prediction error 의 signal"**. 매 dopamine 의 modern view 는 pleasure 의 X, 매 *expected vs actual reward 의 차이* 의 broadcast. 매 Schultz (1997) 의 monkey VTA recording 의 RL 의 TD-error 의 isomorphism 의 establish. 매 product UX, addiction design, RL algorithm 의 shared substrate.
|
||
|
||
## 매 핵심
|
||
|
||
### 매 RPE (Reward Prediction Error)
|
||
- **Positive RPE**: 매 expected 보다 better. 매 dopamine burst.
|
||
- **Zero RPE**: 매 fully predicted. 매 baseline firing.
|
||
- **Negative RPE**: 매 expected 보다 worse. 매 firing dip.
|
||
|
||
### 매 RL 의 TD-error 와 의 mapping
|
||
- 매 δ = r + γV(s') − V(s).
|
||
- 매 dopamine neuron 의 firing rate 의 δ 의 encode (Schultz, Dayan, Montague 1997).
|
||
|
||
### 매 응용
|
||
1. Variable-ratio schedule (slot machine, social media feed) — 매 maximal RPE.
|
||
2. Habit formation (intermittent reward).
|
||
3. Anhedonia / addiction 의 dopaminergic dysregulation.
|
||
4. RL agent design (curiosity, intrinsic motivation).
|
||
|
||
## 💻 패턴
|
||
|
||
### TD-learning (dopamine-analog)
|
||
```python
|
||
import numpy as np
|
||
|
||
def td_update(V, s, r, s_next, alpha=0.1, gamma=0.9):
|
||
"""V: value table. δ = TD error = 'dopamine signal'."""
|
||
delta = r + gamma * V[s_next] - V[s] # ← RPE
|
||
V[s] += alpha * delta
|
||
return delta # log this; it's the 'dopamine'
|
||
|
||
V = np.zeros(10)
|
||
for episode in range(1000):
|
||
s, r, s_next = sample_transition()
|
||
rpe = td_update(V, s, r, s_next)
|
||
```
|
||
|
||
### Curiosity-driven exploration (intrinsic dopamine analog)
|
||
```python
|
||
# Random Network Distillation (Burda 2018)
|
||
class RND(nn.Module):
|
||
def __init__(self):
|
||
super().__init__()
|
||
self.target = nn.Sequential(nn.Linear(64, 128), nn.ReLU(), nn.Linear(128, 64))
|
||
self.predictor = nn.Sequential(nn.Linear(64, 128), nn.ReLU(), nn.Linear(128, 64))
|
||
for p in self.target.parameters(): p.requires_grad = False
|
||
|
||
def intrinsic_reward(self, obs):
|
||
with torch.no_grad():
|
||
target = self.target(obs)
|
||
pred = self.predictor(obs)
|
||
return ((target - pred) ** 2).mean(-1) # novelty bonus
|
||
```
|
||
|
||
### Variable-ratio schedule simulator
|
||
```python
|
||
def variable_ratio_session(p_reward=0.1, n_pulls=100):
|
||
rpe_log = []
|
||
expected = p_reward # learned expectation
|
||
for _ in range(n_pulls):
|
||
r = 1.0 if np.random.rand() < p_reward else 0.0
|
||
rpe = r - expected
|
||
expected += 0.05 * rpe # slow learning
|
||
rpe_log.append(rpe)
|
||
return rpe_log
|
||
# Pattern: high-amplitude RPE persists → "addictive" engagement
|
||
```
|
||
|
||
### Hyperbolic discounting (dopamine-future)
|
||
```python
|
||
def hyperbolic_value(reward, delay, k=0.1):
|
||
"""Real human/animal — closer to hyperbolic than exponential."""
|
||
return reward / (1 + k * delay)
|
||
```
|
||
|
||
### Opponent process (reward + aversion)
|
||
```python
|
||
# Two-system: dopamine (reward) + serotonin (aversion / patience)
|
||
def dual_system_update(V_reward, V_aversion, r_pos, r_neg, s, s_next, alpha=0.1, gamma=0.9):
|
||
delta_reward = r_pos + gamma * V_reward[s_next] - V_reward[s]
|
||
delta_aversion = r_neg + gamma * V_aversion[s_next] - V_aversion[s]
|
||
V_reward[s] += alpha * delta_reward
|
||
V_aversion[s] += alpha * delta_aversion
|
||
return delta_reward, delta_aversion
|
||
```
|
||
|
||
## 매 결정 기준
|
||
|
||
| 상황 | Insight |
|
||
|---|---|
|
||
| Habit-forming product | Variable-ratio reward (Slot machine schedule) |
|
||
| Sustained engagement | Mix predictable + unpredictable wins |
|
||
| Avoid burnout | Avoid pure RPE-maximization (ethical concern) |
|
||
| RL exploration stuck | Add intrinsic reward (RND, ICM) |
|
||
| Anhedonia in user | Reduce expectation, surprise with low-cost wins |
|
||
|
||
**기본값**: 매 RPE-aware design — but 매 ethics 의 weight (manipulation 의 risk).
|
||
|
||
## 🔗 Graph
|
||
- 부모: [[Reinforcement Learning]] · [[Neuroscience]]
|
||
- 변형: [[Serotonin]] · [[Norepinephrine]]
|
||
- 응용: [[Habit Formation]] · [[Game Design]] · [[Recommender Systems]]
|
||
- Adjacent: [[TD-Learning]] · [[Curiosity-Driven RL]] · [[Behavioral Economics]]
|
||
|
||
## 🤖 LLM 활용
|
||
**언제**: 매 product UX 의 retention mechanic 의 audit. 매 dark-pattern 의 detection.
|
||
**언제 X**: 매 clinical advice. 매 LLM 의 medical claim 의 X.
|
||
|
||
## ❌ 안티패턴
|
||
- **Dopamine = pleasure 의 simplification**: 매 X. 매 RPE 의 signal — pleasure 는 separate (opioid).
|
||
- **Pure exploitation (no novelty)**: 매 user 의 RPE 의 0 의 disengage.
|
||
- **Manipulative dark pattern**: 매 ethical violation. 매 design 의 audit 의 mandatory.
|
||
|
||
## 🧪 검증 / 중복
|
||
- Verified (Schultz 1997 Science, Sutton & Barto 2018, Berridge 2007).
|
||
- 신뢰도 A.
|
||
|
||
## 🕓 Changelog
|
||
| 날짜 | 변경 |
|
||
|---|---|
|
||
| 2026-05-08 | Phase 1 |
|
||
| 2026-05-10 | Manual cleanup — RPE / TD-learning isomorphism + UX implication |
|