Training Neural Networks in Cyber-Physical Systems using Design Space Exploration and Co-Simulation

In this paper we present our ongoing research aimed at providing methodologies and tools for reasoning about behavioural properties of AI models in cyber-physical systems. In particular, we envision the use of co-simulation and design space exploration technologies for training ML networks and for validation of the network model for robustness. In detail, we present a simple case study on training a neural network model to approximate the controller action of a thermostat on a room. We show how we can build a training dataset of thousand samples using co-simulation and then using it for training a fully-connected neural network, obtaining ~96% accuracy on testing data. We then show the closed-loop behaviour of the neural model co-simulating it with the controlled system, reporting a mean variation of ∼ 0.3 deg with respect to the nominal controller.