2nd/10_Wiki/Topics/AI_and_ML/AI 추적 논리(AI Pursuit Logic).md

---
id: wiki-2026-0508-ai-추적-논리-ai-pursuit-logic
title: AI Pursuit Logic (AI 추적 논리)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [AI 추적, pursuit logic, aggro, threat detection, leash range, NPC chase]
duplicate_of: none
source_trust_level: B
confidence_score: 0.85
verification_status: conceptual
tags: [game-ai, pursuit, aggro, threat-system, npc-design, baiting, leash, fsm, behavior-tree]
raw_sources: []
last_reinforced: 2026-05-09
github_commit: pending
inferred_by: Claude Opus 4.7 (manual cleanup 2026-05-09)
tech_stack:
  language: game design / Unity / Unreal
  applicable_to: [Game Design, Combat AI, RTS, MMO, FPS]
---

# AI Pursuit Logic (AI 추적 논리)

## 📌 한 줄 통찰 (The Karpathy Summary)
> **"매 enemy 의 어떤 target 의 chase?"**. Aggro / threat / line-of-sight / leash 의 4 component. 매 RTS / MMO / FPS 의 fundamental. **매 deterministic = baiting exploitable, 매 random = unpredictable feel**.

## 📖 구조화된 지식 (Synthesized Content)

### 매 component
1. **Detection**: 매 enemy 의 sight cone / sound / threat radius.
2. **Threat / aggro**: 매 player 의 weight (proximity, damage dealt, special).
3. **Pursuit decision**: 매 stance / state 의 chase / hold.
4. **Leash**: 매 chase 의 max distance.
5. **De-aggro**: 매 reset condition.

### 매 stance / state

#### Active pursuit (Fire at Will / Aggressive)
- 매 enemy detect → chase to engage.
- 매 wide pursuit radius.
- Vulnerable to baiting.

#### Defensive (Hold Position / Stand Ground)
- 매 stay in spot.
- 매 in-range only engage.
- Bait 의 immune.

#### Patrol
- 매 fixed route.
- 매 detect → pursue 일정 time → return.
- 매 player 의 timing exploit.

#### Guarding
- 매 specific target / area.
- 매 leash 의 strict.
- 매 player 의 line-of-sight break 의 reset.

#### Conditional (low HP, ally death)
- 매 trigger 의 stance change.
- Berserk / flee.

### Aggro / threat system

#### Proximity-based (simple)
- 매 distance 의 inverse 의 weight.
- 매 closest = highest priority.

#### Damage-based
- 매 damage dealt 의 cumulative.
- 매 healer / DPS 의 high priority.
- MMO 의 tank role 의 enable.

#### Stealth / detection
- 매 sight cone (FPS, stealth game).
- 매 noise level.
- 매 disguise / blend.

#### Special abilities
- 매 taunt / provoke 의 instant aggro.
- 매 invisibility / cloak 의 break aggro.
- 매 mind control.

### Leash mechanism
- 매 chase 의 max range.
- 매 exceed = full reset (heal, return).
- 매 cheese 의 prevent (kite + reset attack).

```
Chase distance > maxLeash:
  - Stop pursuit.
  - Reset HP to full.
  - Return to spawn.
  - 매 player 의 punish.
```

### Wild goose chase 의 exploit
1. Player unit (fast / cheap) 의 enemy 의 lure.
2. Enemy 의 pursue.
3. Player 의 ambush location.
4. Enemy 의 trap.

→ 매 RTS / MMO 의 universal tactic.

### Modern variants

#### Adaptive aggro
- 매 player behavior 의 learn.
- 매 cheese pattern 의 detect.
- 매 counter-strategy.

#### Squad coordination
- 매 enemy 의 group communicate.
- 매 player 의 line-of-sight 의 share.
- 매 flanking maneuver.

#### Memory / persistence
- 매 enemy 의 last known player position.
- 매 search 의 specific area.
- 매 continued investigation.

#### Investigation state (stealth)
- 매 noise → suspicious → investigate → return / aggro.
- 매 partial detection 의 layered.

### Game examples

#### Classic RTS (StarCraft, AoE, Warcraft)
- A-move = active pursuit.
- Stop / Hold = anchor.
- 매 leash range.

#### MMO (WoW, FFXIV)
- 매 mob 의 leash range (tunable).
- 매 tank 의 threat.
- 매 reset 의 full heal (anti-cheese).

#### Stealth (Splinter Cell, MGS)
- 매 detection meter.
- 매 sound radius.
- 매 alert state cycle.

#### FPS (Halo, Doom)
- 매 enemy 의 cover seeking.
- 매 line-of-sight break 의 search.
- 매 grenade 의 flush.

### Player exploit pattern

#### 1. Bait + lure
- 매 cheap unit 의 sacrifice.
- 매 enemy 의 chase.
- 매 ambush.

#### 2. Kiting
- 매 ranged 의 attack + retreat.
- 매 enemy 의 perpetual pursue.
- 매 attrition kill.

#### 3. Line-of-sight break
- 매 wall 의 hide.
- 매 enemy 의 lose vision.
- 매 reset / stealth.

#### 4. Leash exploit
- 매 enemy 의 leash 끝.
- 매 attack 후 retreat.
- 매 enemy 의 reset (full heal — but separated).

→ 매 MMO 의 banned (exploit) 또는 intended (skill).

#### 5. Aggro management
- 매 player 의 selective threat.
- 매 boss 의 specific target.
- 매 strategic positioning.

## 💻 코드 패턴 (Code Patterns)

### Threat / aggro table
```csharp
public class ThreatTable {
    Dictionary<Player, float> threats = new();
    
    public void AddThreat(Player p, float amount) {
        threats[p] = threats.GetValueOrDefault(p, 0) + amount;
    }
    
    public Player GetTopThreat() {
        if (threats.Count == 0) return null;
        return threats.OrderByDescending(x => x.Value).First().Key;
    }
    
    public void Decay(float deltaTime) {
        foreach (var key in threats.Keys.ToList()) {
            threats[key] *= Mathf.Pow(0.9f, deltaTime);   // 10s half-life
        }
    }
    
    public void OnPlayerDeath(Player p) { threats.Remove(p); }
}
```

### Pursuit FSM
```csharp
public enum AIState { Idle, Patrol, Suspicious, Pursuing, Engaging, Returning }

public class EnemyAI : MonoBehaviour {
    public AIState state = AIState.Idle;
    public Vector3 spawnPoint;
    public float leashRange = 30f;
    public float sightRange = 15f;
    public Transform target;
    
    void Update() {
        switch (state) {
            case AIState.Idle:
                if (DetectPlayer()) state = AIState.Pursuing;
                break;
            
            case AIState.Pursuing:
                if (Vector3.Distance(transform.position, spawnPoint) > leashRange) {
                    state = AIState.Returning;
                    target = null;
                } else if (InAttackRange()) {
                    state = AIState.Engaging;
                } else {
                    MoveToward(target.position);
                }
                break;
            
            case AIState.Engaging:
                Attack(target);
                if (!InAttackRange()) state = AIState.Pursuing;
                break;
            
            case AIState.Returning:
                if (ReachedSpawn()) {
                    HealFully();   // anti-cheese
                    state = AIState.Idle;
                }
                MoveToward(spawnPoint);
                break;
        }
    }
    
    bool DetectPlayer() {
        Player p = FindNearestPlayer();
        if (p == null) return false;
        if (Vector3.Distance(transform.position, p.transform.position) > sightRange) return false;
        if (!HasLineOfSight(p)) return false;
        target = p.transform;
        return true;
    }
}
```

### Stance-based pursuit (RTS)
```csharp
public enum Stance { Passive, Defensive, Aggressive, Hold }

public class RTSUnit : MonoBehaviour {
    public Stance stance = Stance.Defensive;
    public float pursuitRadius = 10f;
    
    void Update() {
        switch (stance) {
            case Stance.Passive:
                // Don't engage. Move only on player command.
                break;
            
            case Stance.Defensive:
                // Engage in attack range only.
                Enemy nearby = FindEnemyInRange(attackRange);
                if (nearby != null) Attack(nearby);
                break;
            
            case Stance.Aggressive:
                // Pursue any enemy in radius.
                Enemy enemy = FindEnemyInRange(pursuitRadius);
                if (enemy != null) {
                    if (InAttackRange(enemy)) Attack(enemy);
                    else MoveToward(enemy.transform.position);
                }
                break;
            
            case Stance.Hold:
                // Like Defensive but no movement.
                Enemy inRange = FindEnemyInRange(attackRange);
                if (inRange != null) Attack(inRange);   // No move.
                break;
        }
    }
}
```

### Sight cone (stealth)
```csharp
public class SightCone : MonoBehaviour {
    public float sightRange = 15f;
    public float sightAngle = 90f;
    public LayerMask obstacleLayer;
    
    public bool CanSee(Transform target) {
        Vector3 dir = (target.position - transform.position).normalized;
        float dist = Vector3.Distance(transform.position, target.position);
        
        if (dist > sightRange) return false;
        if (Vector3.Angle(transform.forward, dir) > sightAngle / 2) return false;
        
        // Line-of-sight check
        if (Physics.Raycast(transform.position, dir, dist, obstacleLayer)) return false;
        
        return true;
    }
}
```

### De-aggro / reset
```csharp
public class DeAggro : MonoBehaviour {
    public float losTimeout = 5f;   // 5s 의 line-of-sight break = reset
    float lastSeenTime;
    
    void Update() {
        if (sight.CanSee(target)) {
            lastSeenTime = Time.time;
        } else if (Time.time - lastSeenTime > losTimeout) {
            ResetAggro();
        }
    }
    
    void ResetAggro() {
        threatTable.Clear();
        target = null;
        state = AIState.Returning;
    }
}
```

### Squad coordination
```csharp
public class Squad : MonoBehaviour {
    List<EnemyAI> members = new();
    Player sharedTarget;
    
    public void OnDetect(Player p) {
        if (sharedTarget == null) {
            sharedTarget = p;
            // Alert all members
            foreach (var m in members) {
                m.target = p.transform;
                m.state = AIState.Pursuing;
            }
        }
    }
    
    public void Coordinate() {
        // 매 member 의 different angle 의 attack.
        for (int i = 0; i < members.Count; i++) {
            float angle = (Mathf.PI * 2 * i) / members.Count;
            Vector3 offset = new Vector3(Mathf.Cos(angle), 0, Mathf.Sin(angle)) * 5;
            members[i].SetMoveTarget(sharedTarget.transform.position + offset);
        }
    }
}
```

### Memory / investigation (stealth)
```csharp
public class GuardAI : MonoBehaviour {
    public AIState state = AIState.Patrol;
    Vector3 lastSeenPosition;
    float investigateTime = 10f;
    
    void Update() {
        if (state == AIState.Patrol) {
            if (DetectPlayer()) {
                state = AIState.Pursuing;
                lastSeenPosition = target.position;
            }
        } else if (state == AIState.Pursuing) {
            if (CanSeePlayer()) {
                lastSeenPosition = target.position;
            } else {
                state = AIState.Investigating;
            }
        } else if (state == AIState.Investigating) {
            MoveToward(lastSeenPosition);
            
            if (ReachedDestination()) {
                if (CanSeePlayer()) {
                    state = AIState.Pursuing;
                } else if (timeInState > investigateTime) {
                    state = AIState.Patrol;   // Give up.
                }
            }
        }
    }
}
```

### Anti-cheese (RTS leash + heal)
```csharp
public class AntiCheese : MonoBehaviour {
    public float maxLeash = 30f;
    Vector3 spawnPoint;
    
    void Update() {
        if (Vector3.Distance(transform.position, spawnPoint) > maxLeash) {
            // 매 player 의 kite + reset attack 의 punish.
            ResetAggro();
            FullHeal();
            ReturnHome();
        }
    }
}
```

## 🤔 의사결정 기준 (Decision Criteria)

| 상황 | 추천 pursuit logic |
|---|---|
| RTS unit | Stance-based (player choice) |
| MMO mob | Threat table + leash |
| FPS enemy | Sight cone + cover seeking |
| Stealth game | Sight + sound + investigation state |
| Boss fight | Phased + threat + special mechanic |
| Roguelike | Simple proximity + variety |

**기본값**: Threat table + leash + LOS-based reset. Anti-cheese.

## ⚠️ 모순 및 업데이트 (Contradictions & Updates)
- **Predictable vs unfair**: 매 deterministic 의 baiting exploit, 매 random 의 frustrating.
- **Aggro 의 transparency**: MMO 의 tank role 의 explicit threat. 매 player 의 strategic.
- **Leash 의 anti-cheese vs immersion**: 매 reset 의 unrealistic feel.
- **AI 의 squad coordination 의 cost**: 매 sophisticated 의 dev expensive.

## 🔗 지식 연결 (Graph)
- 부모: [[Game-AI-Design]] · [[Combat-AI]] · [[NPC-Behavior]]
- 변형: [[Aggro-Threat-System]] · [[Stance-Based-AI]] · [[Sight-Cone-Detection]] · [[Investigation-State]]
- 응용: [[AI-Exploitation]] · [[Baiting-Tactics]] · [[Kiting-Strategy]] · [[Boss-Fight-Design]]
- 매 game: [[War-Commander]] · [[StarCraft-AI]] · [[WoW-Mob-AI]] · [[Splinter-Cell-Stealth]] · [[Halo-Combat-AI]]
- Adjacent: [[Behavior-Tree]] · [[FSM-Game-AI]] · [[Utility-AI]] · [[Squad-Coordination]] · [[Anti-Cheese-Mechanism]]

## 🤖 LLM 활용 힌트 (How to Use This Knowledge)

**언제 이 지식을 쓰는가:**
- 매 game 의 enemy AI design.
- 매 stance / threat system 의 implement.
- 매 stealth game 의 detection.
- 매 boss fight 의 mechanic.
- 매 player exploit 의 review.

**언제 쓰면 안 되는가:**
- Single-player turn-based (different paradigm).
- Walking simulator.
- 매 LLM agent (different domain — partial overlap).

## ❌ 안티패턴 (Anti-Patterns)
- **No leash + open world**: 매 cheese 의 enable.
- **Pure proximity threat**: 매 healer / DPS 의 unfair.
- **No LOS reset**: 매 wall 의 follow 의 unrealistic.
- **No memory**: 매 player 의 hide 후 의 instant forget.
- **No squad coordination**: 매 1 vs 1 의 weak.
- **Fully deterministic**: 매 첫 try 후 trivialize.
- **Excessive randomness**: 매 player 의 frustrating.

## 🧪 검증 상태 (Validation)
- **정보 상태:** verified (concept-level).
- **출처 신뢰도:** B (Game AI Programming Wisdom series, Unity / Unreal documentation, GDC talks).
- **검토 이유:** Manual cleanup. 매 specific game 의 implementation 가 design choice.

## 🧬 중복 검사 (Duplicate Check)
- **기존 유사 문서:** [[AI-Exploitation]] (related), [[Game-AI-Behavior-Tree]] (parent), [[Combat-AI]] (parent).
- **처리 방식:** KEEP (specific focus on pursuit).
- **처리 이유:** Pursuit logic 가 specific component.

## 🕓 변경 이력 (Changelog)
| 날짜 | 변경 내용 | 처리 방식 | 신뢰도 |
|------|-----------|-----------|--------|
| 2026-05-08 | P-Reinforce Phase 1 정규화 | UPDATE | A |
| 2026-05-09 | Manual cleanup — Unity code + 4 component + 매 game example + 안티패턴 추가 | UPDATE | B |