Risk-averse optimal control via homotopy – ROCH

From energy networks to space systems, complex autonomous systems have become pervasive in our society. In this context, the design of increasingly sophisticated methodologies for controlling these systems is of utmost relevance, given that they regularly operate in uncertain and dynamic circumstances. In particular, to mitigate hazardous and possibly catastrophic uncertain perturbations during the decision-making process, one must accurately design and reliably infuse stochastic dynamical models in the control pipeline. Moreover, to optimally balance robustness with respect to the aforementioned perturbations with performance, one must efficiently optimize complex rewards (or costs) online over spaces of controls, i.e., strategies, which are infinite-dimensional. These demanding desiderata call for the design of novel tools for the modelization and optimal control of stochastic systems.
In this project, I will develop and combine original control-theoretic-based learning approaches with novel risk-averse stochastic optimal control techniques to tackle the relevant challenges underneath the modelization and optimal control of stochastic systems. The ultimate objective is to leverage such new methods to devise reliable and scalable algorithms for the efficient and safe-against-uncertainties deployment of autonomous systems in complex uncertain environments. This project is organized into three Work Packages (WP), which respectively aim at proposing reliable learning-based stochastic dynamical models (WP1), efficient resolution of risk-averse stochastic optimal control problems (WP2), and leveraging these scientific achievements to solve challenging application-related problems in space robotics and energy (WP3).