ExAMPC: the Data-Driven Explainable and Approximate NMPC with Physical Insights

ExAMPC makes complex AI controllers more transparent: The methodology uses explainable AI techniques (SHAP and Symbolic Regression) to reveal why advanced controllers like NMPC make specific decisions and how they respond to different physical variables, addressing a key barrier to adopting AI in safety-critical control systems.

Computational efficiency is dramatically improved without sacrificing performance: By reducing trajectory data dimensionality by over 95% while maintaining accuracy, ExAMPC makes advanced control methods practical for real-time applications like autonomous vehicles, where computational speed directly impacts safety.

Physical constraints are better respected: The approach reduces predicted trajectory violations by 93% by incorporating physics-inspired penalties, meaning the AI controller's decisions align more reliably with real-world physical laws—critical for trustworthy autonomous systems.

Predictability of system behavior increases: ExAMPC can accurately forecast the controller's computational requirements and explain worst-case scenarios, helping engineers and operators understand when and why the system might struggle—essential information for responsible deployment in safety-critical applications.

Bridges the gap between classical and AI-based control: By combining the transparency of traditional control methods with data-driven AI insights, ExAMPC offers a middle path that maintains the deterministic, understandable nature engineers trust while gaining AI's performance benefits.