Real-Time Systems Computer Networks and Communications

On-time sensitivity optimization and failure-aware AI orchestration via model predictive control

Alkaraawi, Majid Rahman Lafta, Shengwu, Xiong, Neamah, Osamah N.
Springer Nature · 2026-09-23 · DOI 10.1007/s11241-026-09459-x

Abstract

Artificial Intelligence (AI) models are typically evaluated using diagnostic performance metrics and offline inference time. However, queuing under variable workload may cause decisions to exceed operational deadlines. This study formulates AI-model orchestration as a constrained online control problem. The study is based on Monte Carlo simulation results for evaluating Model Predictive Control (MPC) policies with baseline policies. Decision-making depends on the controller’s selection of a candidate model based on several factors: observed workload state, forecast arrivals, model-dependent service times, deadline feasibility, and GPU-memory constraints. The framework combines model sensitivity and specificity with simulated deadline compliance through two proposed measures: On-time Sensitivity and Time-Weighted Failure Mode Risk (TW-FMR). Performance evaluation was conducted on two sets of candidate models: literature-derived model set and synthetic model set. In the literature-derived model set, the MPC policy achieved only minor performance differences compared to baseline policies due to similar performance characteristics. In contrast, the synthetic models showed a clear advantage for the MPC policy. For the literature-derived candidate-model set, the simulated MPC policy achieved an On-time Sensitivity of 0.767 and a TW-FMR of 0.0832, while in the synthetic models, the On-time Sensitivity was 0.934, the TW-FMR was 0.0259, and the On-time Rate was 0.978. These results represent the behavior of the proposed orchestration framework under typical assumptions and simulated workloads and do not prove clinical efficacy or deployment safety.

Read at publisher →