2nd/10_Wiki/Topics/AI_and_ML/Dynamic Difficulty Adjustment (DDA).md

---
id: wiki-2026-0508-dda
title: Dynamic Difficulty Adjustment (DDA)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [DDA, dynamic difficulty, rubber banding, AI Director, flow state, adaptive difficulty]
duplicate_of: none
source_trust_level: B
confidence_score: 0.85
verification_status: applied
tags: [game-design, dda, dynamic-difficulty, flow, ai-director, player-modeling, adaptive]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: game design
  applicable_to: [Game Design, Adaptive Learning, ML Curriculum]
---

# Dynamic Difficulty Adjustment (DDA)

## 매 한 줄
> **"매 player skill 의 real-time 매 difficulty 의 dial"**. Csikszentmihalyi 의 Flow zone 의 maintain. 매 Left 4 Dead 의 AI Director, 매 racing 의 rubber banding. 매 modern: 매 ML-driven player model + 매 ethical detect (gaming the system). 매 LMS adaptive learning 의 same principle.

## 매 핵심

### 매 mechanism
1. **Player metric**: 매 win rate, HP remaining, time, deaths.
2. **Skill estimate**: 매 sliding window of recent.
3. **Adjustment**: 매 enemy stats / spawn / hint.
4. **Subtle**: 매 obvious 의 player 의 frustrate.

### 매 famous example
- **Left 4 Dead AI Director** (Valve): 매 zombie spawn 의 control.
- **Resident Evil 4**: 매 difficulty 의 silent.
- **Mario Kart**: 매 rubber banding (item, speed).
- **Crash Bandicoot**: 매 attempt 후 의 adjust.
- **Mario AI** (research): 매 student 의 player.

### 매 method
- **Heuristic**: 매 simple rule.
- **Bayesian player model**: 매 skill posterior.
- **RL-based**: 매 difficulty 의 policy 학습.
- **MCTS-based**: 매 lookahead.

### 매 ethical / design constraint
- **Subtle**: 매 player 의 perception X.
- **Fair feel**: 매 effort-reward.
- **Not patronize**: 매 매 player 의 skill ↑ 의 acknowledge.
- **No gaming detection**: 매 player 의 intentional poor 의 abuse.

### 매 응용
1. **Action / shooter**: 매 enemy difficulty.
2. **Racing**: 매 rubber banding.
3. **Puzzle**: 매 hint.
4. **MMO raid**: 매 boss adjustment.
5. **Edu game / LMS**: 매 question difficulty.
6. **AI tutor**: 매 explanation depth.
7. **Aim training**: 매 [[Cognitive Training Software (eg Aim Lab_KovaaKs)]] 의 adaptive scenario.

### 매 modern AI 응용
- **Player modeling**: 매 ML 의 skill 의 estimate.
- **RL-based DDA**: 매 difficulty controller 의 train.
- **Generative content**: 매 procedurally adapted level.

## 💻 패턴

### Sliding window skill estimator
```ts
class SkillEstimator {
  private history: number[] = [];
  
  recordOutcome(win: boolean, time: number, hpLost: number) {
    const score = (win ? 1 : 0) * (60 / Math.max(time, 1)) * (1 - hpLost / 100);
    this.history.push(score);
    if (this.history.length > 20) this.history.shift();
  }
  
  estimateSkill(): number {
    if (!this.history.length) return 0.5;
    return this.history.reduce((a, b) => a + b, 0) / this.history.length;
  }
}
```

### AI Director-style (probabilistic)
```ts
class AIDirector {
  threatLevel = 0;  // 0-1
  
  update(player: Player) {
    const recentDamage = player.recentDamageTaken();
    const ammoLow = player.ammo < 30;
    const downtime = player.timeWithoutCombat();
    
    // 매 raise threat 의 calm
    if (downtime > 30) this.threatLevel = Math.min(1, this.threatLevel + 0.05);
    
    // 매 ease 의 stress
    if (recentDamage > 50 || ammoLow) this.threatLevel = Math.max(0, this.threatLevel - 0.1);
  }
  
  shouldSpawnEnemy(): boolean {
    return Math.random() < this.threatLevel;
  }
  
  enemyComposition() {
    if (this.threatLevel < 0.3) return 'easy_pack';
    if (this.threatLevel < 0.7) return 'mixed';
    return 'horde';
  }
}
```

### Rubber banding (racing)
```ts
function applyRubberBand(car: Car, leader: Car) {
  const distanceBehind = leader.position - car.position;
  
  // 매 subtle catch-up
  if (distanceBehind > 100) {
    car.maxSpeed *= 1.05;  // 매 5% boost
    car.itemDropChance *= 1.5;  // 매 better items
  } else if (distanceBehind < -50) {
    // 매 leader 의 small handicap
    car.maxSpeed *= 0.98;
  }
}
```

### Bayesian player model (IRT-inspired)
```python
import numpy as np

class BayesianPlayer:
    def __init__(self, prior_mean=0, prior_var=1):
        self.skill_mean = prior_mean
        self.skill_var = prior_var
    
    def update(self, item_difficulty, success):
        """매 item response theory."""
        # 매 expected probability
        p_success = 1 / (1 + np.exp(-(self.skill_mean - item_difficulty)))
        
        # 매 posterior update (simplified Kalman)
        info = p_success * (1 - p_success)
        self.skill_var = 1 / (1 / self.skill_var + info)
        self.skill_mean += self.skill_var * info * (int(success) - p_success)
    
    def select_difficulty(self, items, target_p=0.7):
        """매 just-beyond comfort."""
        return min(items, key=lambda d: abs(
            1 / (1 + np.exp(-(self.skill_mean - d))) - target_p
        ))
```

### RL-based DDA
```python
class DifficultyController:
    """매 RL 의 매 player 의 engagement 의 maximize."""
    
    def __init__(self):
        self.policy = QNetwork()
    
    def select_difficulty(self, player_state):
        actions = ['decrease', 'maintain', 'increase']
        q_values = self.policy(player_state)
        return actions[q_values.argmax()]
    
    def reward(self, player_state):
        # 매 engagement = match length + retention + flow indicator
        return (
            player_state.session_length * 0.3 +
            (1 if player_state.continued_after_match else 0) * 0.5 +
            player_state.subjective_satisfaction * 0.2
        )
```

### Detect gaming the system
```python
def detect_throw_match(player_history):
    """매 player 의 intentional poor 의 detect."""
    recent = player_history[-10:]
    
    # 매 sudden drop in performance
    historical_avg = np.mean([h.score for h in player_history[:-10]])
    recent_avg = np.mean([h.score for h in recent])
    
    if recent_avg < historical_avg * 0.4:
        return 'WARN: possible throwing for DDA exploit'
    return None
```

### Subtle communication (UX)
```ts
function hidePlayerSeesAdjustment() {
  // 매 ❌ 매 obvious
  if (showText) showMessage('Difficulty decreased due to your performance');
  
  // 매 ✅ 매 silent
  // 매 just adjust enemy stats internally.
  // 매 player 의 say "I'm getting better!" 매 actually 의 difficulty 의 ease.
}
```

### A/B test (DDA effectiveness)
```python
def ab_test_dda(players_a_no_dda, players_b_with_dda):
    return {
        'session_length_a': mean(p.session_length for p in players_a_no_dda),
        'session_length_b': mean(p.session_length for p in players_b_with_dda),
        'retention_d7_a': retention(players_a_no_dda, days=7),
        'retention_d7_b': retention(players_b_with_dda, days=7),
        'satisfaction_a': mean(p.satisfaction for p in players_a_no_dda),
        'satisfaction_b': mean(p.satisfaction for p in players_b_with_dda),
    }
```

### LMS adaptive (similar pattern)
```python
def adaptive_quiz(student, item_pool):
    skill = student.estimated_skill
    
    # 매 zone of proximal development
    target_difficulty = skill + 0.5  # 매 just-beyond
    next_item = min(item_pool, key=lambda i: abs(i.difficulty - target_difficulty))
    
    response = present(next_item, student)
    student.update_skill(next_item.difficulty, response.correct)
    return next_item
```

### Procedural difficulty (level generation)
```python
def generate_level(player_skill):
    return {
        'enemy_count': int(10 + player_skill * 20),
        'enemy_hp': 100 * (1 + player_skill * 0.5),
        'puzzle_complexity': int(player_skill * 5),
        'time_limit': 300 - int(player_skill * 100),
        'powerup_density': max(0.1, 0.5 - player_skill * 0.3),
    }
```

## 매 결정 기준
| Genre | DDA approach |
|---|---|
| Action / Shooter | AI Director (Valve-style) |
| Racing | Rubber banding (subtle) |
| Puzzle | Hint frequency |
| MMO raid | Stat-tier difficulty |
| RPG | Side content |
| Edu game | Adaptive item (IRT) |
| Esports / competitive | NO DDA (fairness) |
| Casual mobile | Subtle DDA + retention focus |

**기본값**: 매 Bayesian player model + 매 subtle adjust + 매 A/B test + 매 detect gaming.

## 🔗 Graph
- 부모: [[Game-Design]] · [[Adaptive-Learning]]
- 변형: [[AI-Director]] · [[Rubber-Banding]] · [[Player-Modeling]]
- 응용: [[Cognitive Training Software (eg Aim Lab_KovaaKs)]] · [[Corporate-LMS-Training]] (adaptive) · [[Cognitive-Evaluation-Theory]]
- Adjacent: [[Default Mode Network (DMN)]] (flow) · [[Deliberate-Practice]] · [[Combined Arms (제병협동) 전술]]

## 🤖 LLM 활용
**언제**: 매 game design. 매 adaptive learning system. 매 personalized challenge.
**언제 X**: 매 esports / fairness-critical (no DDA).

## ❌ 안티패턴
- **Obvious DDA**: 매 player 의 perception → 매 motivation lose.
- **Patronizing**: 매 매 player 의 skill 의 acknowledge X.
- **Punishing skilled play**: 매 better → 매 harder feel of unfair.
- **No detection 의 throw**: 매 exploit.
- **DDA in competitive**: 매 fairness violation.
- **Too aggressive adjust**: 매 whiplash.

## 🧪 검증 / 중복
- Verified (Csikszentmihalyi Flow, Valve AI Director paper, Hunicke MDA).
- 신뢰도 B.
- Related: [[Cognitive Training Software (eg Aim Lab_KovaaKs)]] · [[Corporate-LMS-Training]] · [[Cognitive-Evaluation-Theory]] · [[Deliberate-Practice]] · [[Combined Arms (제병협동) 전술]].

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — mechanism + 매 AI Director / rubber band / Bayesian / RL / detect-throw code |