Active Inference for Digital Twins: Predicting and Optimizing IoT Processing Service Performance

Abstract

Digital Twins (DTs) are emerging as enablers for real-time monitoring and control in IoT systems. In this work, we propose a DT enabled with Active Inference (AIF), a framework rooted in the Free Energy Principle, to endow DTs with predictive and decision-making capabilities. Our approach is evaluated on a realistic edge computing scenario where two co-located video processing services—a computer vision pipeline using YOLOv8 and a QR code reader—compete for limited CPU resources. The DT continuously infers physical twin state, predicts future performance, and selects actions to meet Service Level Objectives (SLOs). In a series of experiments, we validate the predictive accuracy and control effectiveness of the AIF-enabled DT. Notably, the DT achieves a cumulative SLO fulfillment above 80% for both services, and predicted throughput trajectories show high alignment with real observations, confirmed through statistical testing.