PhD student
Team : MLIA
Arrival date : 11/01/2021
    Sorbonne Université - LIP6
    Boîte courrier 169
    Couloir 25-26, Étage 5, Bureau 520
    4 place Jussieu
    75252 PARIS CEDEX 05

Tel: +33 1 44 27 71 33, Lucas.Schott (at)

Supervision : Sylvain LAMPRIER

Co-supervision : HAJRI Hatem (System X)

Robust Reinforcement Learning

Deep Reinforcement learning (RL) is an area of machine learning for autonomous decisional agent using neural netwroks. The agent learns in an environment in which it gets observations, takes decisions and receives rewards or penalties. Agents learnt from RL can perform extremely well under the conditions in which they are learnt, but an important problem in RL is to obtain agents that are robust to perturbations in the observations and in the environment. The thesis work will focus on : - Study the robustness of neural networks of RL agents to perturbations in the inputs. - Apply certifiable defence techniques and random smoothing to RL neural networks and study the impact on the robustness of the resulting policy. - Determine guaranteed robustness bounds. - Study the applications of these approaches in various AI applications such as navigation, control, natural language processing, etc