Robust control-driven counterfactual generation for uncertain systems

This work is devoted to the discussion of a robust approach for counterfactual generation in a control framework. Counterfactual is a concept from the field of logic, which in last years has been adopted in the context of artificial intelligence to describe the minimum changes in the input variables required to vary the outcome of a classification algorithm. In general, this framework for counterfactuals shows some limitations inherently due to the fact the most of the machine learning models leverage black box approaches, overlooking the physics of the underlying system. In this work we discuss a control system approach to derive counterfactuals that are informed by the knowledge about the system dynamics, in the very general case of a system affected by model parameter uncertainties.