Identification of underlying incentives using inverse optimal control

Dynamical systems are essential for modeling and analysis in many different areas of science and engineering. Observed behavior can often be modeled as dynamical systems that operate by minimizing cost or maximizing reward. This type of behavior is known as optimal control.

However, in applications of optimal control, it is often difficult to design the cost function since it must be carefully selected and tuned to induce an appropriate control response.

This project will develop novel methods that can overcome this problem, and that can be used to understand observed behavior.

The aim is to develop algorithms, for both discrete-time and continuous-time systems, to learn the underlying cost function for which the observed behavior is optimal.

In the discrete-time setting, this will be done by combining our new results on statistically consistent estimators for linear-quadratic inverse optimal control with Koopman operator theory.

In the continuous-time case, this will be done by combining our results on entropy-regularized optimal transport with new insights into how this connects both to optimal control and to state estimation and filtering.

The research will be carried out by myself, as well as by a Postdoc who will be hired in the project, in collaboration with international specialists with relevant competences from my network.