Model validity and tolerance quantification for real-time adaptive approximation

Abstract

Designing a Cyber-physical system (CPS) including modeling the control components and services is a challenging issue. Models and simulations at run-time play a crucial role to implement these control and prediction components. Real-time constraints raise the complexity of designing an efficient CPS system. Having detailed models in making decisions and/or numerous predictions in different contexts is computationally expensive and difficult to schedule on the computational infrastructure. Inspired by substitutability, one strategy for dealing with complex CPS and the contradiction of better real-time performance and reduced cost in CPS is to employ approximated models and switch to the most suited model adaptively at run-time. However, using an approximate model raises the uncertainty on the model's predictions. Nonetheless, the model is appropriate when the uncertainty is within bound. This bound is defined as tolerance which is the permitted amount of uncertainty. In this paper, we propose a method for quantifying the tolerance of cyber-physical systems, where we can switch between the original model and approximated models and how to identify more appropriate models.