Optimal control algorithms for stochastic dynamical systems often rely on sample paths to evaluate performance metrics. To synthesize accurate controls we typically need to generate a large number of paths, which rapidly leads to computationally intractable problems. In this talk, I will present a new algorithm we recently developed for optimal control of nonlinear systems under uncertainty. The algorithm has extremely low memory requirements and it leverages on interior point optimization methods, reactive programming, common subexpression elimination and exact gradients obtained via computational graphs. These features allow us to process simultaneously a very large number of sample paths, and compute open-loop controls over long time horizons. I will also discuss verification and validation of optimal controls using the extended version of the Pontryagin minimum principle. Throughout the lecture I will provide numerical examples and applications of the new algorithm to stochastic path planning problems involving models of unmanned aerial and ground vehicles, and to the Kuramoto-Sivashinsky equation.
Friday, May 10, 2019 - 2:00pm
UC Santa Cruz