🧠 Ensemble_Strategy

Ensemble Strategy: Two AIs Are Better Than One

Washin Village AI Director Tech Notes #4

---

🎯 What is Ensemble?

Just like human teamwork, Ensemble Learning makes multiple AI models work together, combining their judgments for more accurate results.

Core concept: Two heads are better than one.

---

🔍 Why Do We Need Ensemble?

Single model limitations:

Model	Pros	Cons
Unified_v18	High accuracy (79.5%)	Prone to errors on certain categories
Inc_v201	Different training data	Lower accuracy (46%)

However! The probability of both models making the same mistake is very low.

---

💡 Ensemble Strategies

Strategy 1: Voting Mechanism

def ensemble_vote(predictions):
    """Majority voting"""
    from collections import Counter
    votes = Counter([p['class'] for p in predictions])
    return votes.most_common(1)[0][0]

Strategy 2: Weighted Confidence

def ensemble_weighted(predictions, weights):
    """Weighted average confidence"""
    combined = {}
    for pred, weight in zip(predictions, weights):
        for cls, conf in pred.items():
            combined[cls] = combined.get(cls, 0) + conf * weight
    return max(combined, key=combined.get)

Strategy 3: Validation Mode

def ensemble_validate(primary, secondary, threshold=0.85):
    """Primary model predicts, secondary validates"""
    if primary['confidence'] >= threshold:
        return primary['class']

    if primary['class'] == secondary['class']:
        return primary['class']

    return "needs_human_review"

---

📊 Our Configuration

Dual-Model Ensemble

Input Image
    │
    ├─→ Primary Model (Unified_v18) ──→ Prediction + Confidence
    │
    └─→ Validation Model (Inc_v201) ──→ Prediction + Confidence
    │
    ↓
Ensemble Decision Engine
    │
    ↓
Final Result

Decision Rules

Scenario	Action
Primary confidence ≥ 90%	Use primary result directly
Both models agree	Use that result
Models disagree	Choose higher confidence
Both uncertain	Mark for human review

---

📈 Performance Comparison

Configuration	Top-1 Accuracy	Human Review Rate
Primary only	79.5%	0%
Ensemble (voting)	81.2%	0%
Ensemble (validation)	82.8%	5%

Conclusion: Ensemble + validation mode works best!

---

🔧 Actual Code

class EnsembleValidator:
    def __init__(self):
        self.primary = YOLO('Unified_v18.pt')
        self.secondary = YOLO('Inc_v201.pt')

    def predict(self, image):
        # Primary model prediction
        p1 = self.primary.predict(image)

        # High confidence: return directly
        if p1.confidence >= 0.90:
            return p1.class_name, "high_confidence"

        # Secondary model validation
        p2 = self.secondary.predict(image)

        if p1.class_name == p2.class_name:
            return p1.class_name, "validated"

        # Return higher confidence
        if p1.confidence > p2.confidence:
            return p1.class_name, "primary_higher"
        else:
            return p2.class_name, "secondary_higher"

---

💡 Lessons Learned

1. Model diversity matters: Use different architectures or training data
2. Weight allocation: Higher accuracy models get higher weights
3. Speed tradeoffs: Ensemble is slower than single model
4. Threshold tuning: Adjust confidence thresholds based on actual needs

---

Washin Village 🏡 by AI Director