[PyTokMHD] Policy interpretability and physics-informed analysis

[callme-YZ]

Problem

RL policy is black-box

For v3.0 competitive positioning:

• What did RL learn?
• Is it physics-consistent?
• Trust from physics community ⚠️

Current state:

• ✅ Policy trains and works
• ❌ Learned strategy: Unknown
• ❌ Physics interpretation: None

Why Critical

Physics community skepticism:

• "Black-box ML doesn't understand physics"
• Need to show learned control is physics-informed
• Explainability = acceptance

Scientific value:

• Discover new control strategies
• Validate physics intuition
• Gain insights for theory

Proposed Analysis

1. Policy visualization

What does policy do?

• Action distribution analysis
• State-action mapping visualization
• Attention/importance weights (if applicable)

Tools:

• Saliency maps
• SHAP values
• Activation visualization

2. Physics-informed diagnostics

Check physics consistency:

• Does policy respect conservation laws?
• Does it exploit known physics?
• Compare: RL strategy vs physics intuition

Metrics:

• Energy injection timing
• Mode interaction patterns
• Control phase relationship

3. Learned control strategy extraction

Distill policy to rules:

• Decision tree approximation
• Symbolic regression
• Physics-based feature importance

Goal: Extract interpretable control law

4. Comparison with theory

Validate against physics:

• Does RL rediscover known strategies?
• Does it find novel approaches?
• Physics explanation of learned behavior

Implementation

Analysis tools:

• Visualization dashboard
• Physics diagnostic suite
• Automated reporting

Deliverables:

• Interpretability report
• Physics analysis of learned policy
• Comparison: RL vs theory
• Publication figures

Success Criteria

• ✅ Learned strategy interpretable
• ✅ Physics-consistent behavior demonstrated
• ✅ Novel insights discovered (or validated known)

Priority

P2-medium 🟡 - Important for acceptance

Rationale: Turns black-box into physics-informed discovery tool

---

Reported by: 小P ⚛️
Date: 2026-03-23
Context: v3.0 competitiveness gap