Consortium. Riccardo Bonalli (Project Investigator), CNRS and Université Paris-Saclay; Brandon Amos; Alessio Iovine, CNRS and Université Paris-Saclay; Marco Pavone, Stanford University.
Goal. This ANR JCJC project aims at developing both original reinforcement learning techniques to design reliable control models and novel stochastic optimal control tools to tackle non-convex risk-averse stochastic optimal control problems. The ultimate objective is to leverage and combine such new methods to devise a reliable and scalable algorithm called ROCH, for efficient and safe-against-uncertainty deployment of autonomous systems.
New open position! Call for one postoctoral position in stochastic analysis and statistical estimation
Topic. The successful candidate will work in the framework of the ANR project ROCH (Risk-averse Optimal Control via Homotopy), directed by Dr Riccardo Bonalli. ROCH aims at developing both original estimation techniques to design reliable models for the dynamics of modern autonomous systems, as well as novel stochastic optimal control tools to successfully steer such models efficiently and safely against complex uncertainties. The postdoc will focus on developing estimation techniques to design reliable models for sophisticated uncertain dynamics. Specifically, the successful candidate will explore one (or more) of the following research directions.
1 - Devise novel techniques to perform non-parametric estimation for SDEs with irregular coefficients driven by Levy-type noises, thus possibly with jumps. When considering particularly regular coefficients and Brownian noise, RKHS-based estimators have been recently combined with the Fokker-Planck formalism to achieve non-asymptotic rates of convergence, e.g., [J6,J9]. The postdoc could investigate how to leverage the benefits of this approach, with the objective of deriving novel non-asymptotic rates of convergence in case of irregular, e.g., Cα, coefficients, first with Brownian noise and then ideally α-stable noises.
2 - Alternatively, the postdoc could investigate how to leverage the aforementioned RKHS-based estimators to directly learn control actions that enable steering SDEs within unexplored regions, without learning the full dynamics in the first place. By enhancing the theory of SDEs with irregular coefficients driven by Levy-type noises, the objective will consist of deriving novel non-asymptotic rates of convergence for RKHS-based estimators that yield safe exploration - e.g., with no collisions - of unexplored regions with high probability.
3 – Finally, for a system of N interacting stochastic particles, the postdoc could explore the possibility of adding a controlled drift to the dynamics of the particles - designed to reduce the probability of finding two or more particles in the same neighbourhood and in case of attractive interactions - prevent collisions between particles from happening. One key goal would therefore be to consider the large-N-limit.
About the position. The position is for 18 months, with a gross monthly salary range of €2,500-€3,300. Starting date at the earliest convenience of the successful candidate, as of the 1st of March 2025 and before the end of the year 2025. The postdoc will join L2S laboratory at Université Paris-Saclay and will be co-advised by Riccardo Bonalli (L2S, Université Paris-Saclay) and Alexandre Richard (MICS, Université Paris-Saclay).
How to apply. Applications should include a CV, a list of publications, a research statement and 2 letters of recommendation, to be sent to Riccardo Bonalli (riccardo.bonalli@centralesupelec.fr) and Alexandre Richard (alexandre.richard@centralesupelec.fr).
Collaborators.
Publications.