Explaining Process Control Optimisation Recommendations via GradientSHAP and Implicit Differentiation

arXiv cs.AI · 6 hours ago

Researchers developed a method combining Implicit Function Theorem sensitivity analysis with SHAP and large language models to explain automated optimisation recommendations in industrial processes. The approach achieved 40 times faster SHAP computation than standard methods while maintaining correlation above 0.99 with traditional KernelSHAP on an industrial grinding mill control problem with 22 input features. The technique enables real-time natural language explanations for operators, addressing the trust gap between algorithm designers and those implementing optimisation recommendations.