Post-Doctoral Research Visit F/M Safe Planning for Markov Decision Processes via Shielding

Inria Rennes Rennes, France

About the position

Context

The post-doc position is part of a collaboration between Inria and Mitsubishi Electric R&D Centre Europe (MERCE) within the FRAIME project on artificial intelligence and formal methods. The project explores, on the one hand, how Formal Methods can provide guarantees on AI systems, and on the other hand how AI can help Formal Methods to be more efficient and easier to use by practitioners. The vision is to intertwine Formal Methods and AI to efficiently design safe systems.
This is a postdoctoral position in the fields of AI planning, formal methods, and reinforcement learning (RL). The position is initially funded for 12 months, but it is possibly extendible.
While this is an academic position based at Inria Rennes, the candidate will collaborate with researchers from both Inria and MERCE, thus benefiting from both academic and industrial research environments.
The work will be supervised by Ocan Sankur (Inria - IRISA https://devine.inria.fr/), Nicolas Waldburger (MERCE). We will moreover collaborate with Francesco Belardinelli from Imperial College London.

Assignment

The main objective is to develop safe planning algorithms with stronger guarantees in Markov decision processes. More precisely, we will study shielding techniques to ensure the safety of policies computed by reinforcement learning (RL) algorithms.
One of the possible object of study is the beyond-worst case framework [1]. Here, algorithms are given to optimize a given objective in expectation while ensuring guarantees in the worst-case. This can be interpreted as a constrained MDP where the constraints are to be satisfied in the worst case. We will revisit this framework with a shielding point of view: the average case can be optimized using RL algorithms while the worst-case guarantees can be ensured by solving two-player zero-sum games.
The work will consist in identifying appropriate applications, building the right setting to apply shielding, which includes pre-shields and post-shields for RL as in [2], and experimentally evaluating the overall approach. Expected theoretical developments concern scalability issues addressed using game abstractions as in [3].
References:
[1] Bruyère, Véronique, et al. "Meet your expectations with guarantees: Beyond worst-case synthesis in quantitative games."
Information and Computation
254 (2017): 259-295
[2] Anand, Ashwani, et al. "Follow the STARs: Dynamic $\omega $-Regular Shielding of Learned Policies." arXiv preprint arXiv:2505.14689
(2025).
[3] Luca de Alfaro, Patrice Godefroid, and Radha Jagadeesan. 2004. Three-Valued Abstractions of Games: Uncertainty, but with Precision. In Proceedings of the 19th Annual IEEE Symposium on Logic in Computer Science (LICS '04). IEEE Computer Society, USA, 170–179.

Main activities

This is a fully academic post-doc position. The candidate is expected to conduct research and work on applications in collaboration with other researchers, write papers, and present their research in conferences.

Skills

PhD in computer science
An excellent track of publications: the candidate must have published theoretical papers in good conferences and journals, including mathematical proofs
Ability to write rigorous mathematical proofs in full autonomy
Background in probability, Markov chains, MDPs
Knowledge about reinforcement learning and planning are a plus but not necessary for candidates with a strong theoretical background on MDPs.
Interest in applications of planning and formal methods
The candidate is expected to take initiatives such as reading about the state of the art, and coming up with research ideas
Good level of English
Good communication and reporting skills
An interest in collaborative work

This listing was collected from a public source and is reproduced here for information only. Always confirm the details on the original posting before applying.
View the original posting

Similar positions