AS Seminar: On the synthesis and crowdsourcing of control policies for autonomous agents from data

Control design often relies on mathematical models. Unfortunately, detailed models are not always available and/or can be hard to identify. In this context, we present some recent results enabling the synthesis of policies directly from noisy data for (possibly, nonlinear) systems affected by actuation constraints. This is done by recasting the design problem as a (convex) optimal data-driven control problem. By leveraging the probabilistic approach, we give an explicit algorithmic solution to the control problem. Then, after discussing the results, we investigate the question of whether autonomous agents can find a solution to the problem by crowdsourcing knowledge from others, rather than solving it from scratch. We term this (non-convex) problem as the crowdsourcing problem and present an algorithmic procedure that gives an explicit approximate solution to the problem. The theoretical results are illustrated via numerical examples from applications involving autonomous and connected vehicles.