Bridging Predictive ML and Prescriptive Operations Research: Resource Allocation via Mamba-Based Model Prediction Control

Integrating machine learning (ML) into operational digital twins is challenging as demand often changes over time, while evaluating resource decisions requires computationally expensive, and often non-differentiable simulation. We present SMPC-RA, a State-space Model Predictive Control (MPC) for Resource Allocation that bridges ML and operations research (OR). A differentiable Mamba-based surrogate maps demand logs and resource configurations directly to key performance indicators (KPIs), replacing the simulator during planning. An offset-free correction layer adapts predictions online using a Kalman-tracked disturbance and configuration-specific kernel memory, facilitating KPI improvements without retraining the surrogate. We demonstrate the proposed framework with data from a job shop digital twin testbed.