Fidelity-Aware Inference Services in DT-Assisted Edge Computing via Service Model Retraining

The Digital Twin (DT) technique enables seamless integrations between the physical and virtual worlds. By continuously synchronizing DTs with their physical counterparts, DTs can provide accurate reflections of physical objects and facilitate high-fidelity inference services based on service models. Orthogonal to the DT technology, Mobile Edge Computing (MEC) has been envisioning as a promising paradigm for providing intelligent services to users while meeting stringent delay and accuracy requirements. In this paper, we investigate fidelity-aware inference services in a DT-assisted MEC network where there are multiple source DTs providing new updated training data to service models often. We jointly schedule mobile devices to upload their update data to their DTs, and choose service models for retraining using their updated source DT data over a given time horizon. We further assume that the previous version of each service model can still serve its users during its retraining period, while a retrained service model can provide high-fidelity services to its users. To this end, we first formulate two novel optimization problems: the model instance placement problem that assigns model instances to cloudlets in an MEC network so that the total placement cost of all service models is minimized, and the cumulative utility maximization problem to maximize the cumulative fidelity of all service models over a given time horizon, by jointly scheduling mobile devices to upload their update data to their DTs and service models to be trained using their updated source DT data at each time slot. We then formulate an integer linear programming (ILP) solution for the model instance placement problem when the problem size is small; otherwise we develop an approximate solution to the problem, at the expense of moderate resource violations. We also devise an efficient online algorithm for the cumulative utility maximization problem. We finally evaluate the performance of the proposed algorithms via simulations, and the simulation results demonstrate that the proposed algorithms are promising.