Simulation cannot explain CPS behaviour without the hidden influences

The nastiest trap in simulating cyber-physical systems is taking an incomplete model for an explanation of real behaviour. If an important interaction with the environment is not described, the experiment's results may be correct inside the simulation yet insufficient for understanding the system itself.

In a CPS the artefacts of different teams usually converge: some describe control, others mechanics, others the environment. Behaviour arises from their interaction, including links that do not reduce to reading sensors or driving actuators directly.

The authors propose describing such links through the concept of Influences. The idea is not to build a perfect digital copy of the world at once but to widen and refine simulation campaigns as new results arrive: find gaps in the model and probe them with the next experiments.

The approach is demonstrated on a mobile robot in co-simulation with Simulink/Gazebo. The practical value here lies precisely in organising the search for unknown factors: simulation becomes not a final verdict but a tool for successively refining hypotheses about the system's behaviour.

But there is an important limitation. This is a conceptual framework demonstrated on a single case study; the source offers no quantitative comparison with other methods and does not show how much faster or more accurately it helps find hidden interactions. So Influences is worth treating as a way of thinking about a model's boundaries rather than a ready guarantee of a trustworthy simulation.

If you are building a model of a complex system, at what point does a simulation result stop being proof for you and become a reason to look for an unaccounted influence? Source: cs.AI updates on arXiv.org