Distillation of RL Policies with Formal Guarantees via Variational Abstraction of Markov Decision

Florent Delgrange, Ann Nowe, Guillermo A. Pérez

Research output: Unpublished contribution to conferencePoster


We consider the challenge of policy simplification and verification in the context of policies learned through reinforcement learning (RL) in continuous environments. In well-behaved settings, RL algorithms have convergence guarantees in the limit. While these guarantees are valuable, they are insufficient for safety-critical applications. Furthermore, they are lost when applying advanced techniques such as deep-RL. To recover guarantees when applying advanced RL algorithms to more complex environments with (i) reachability, (ii) safety-constrained reachability, or (iii) discounted-reward objectives, we build upon the DeepMDP framework to derive new bisimulation bounds between the unknown environment and a learned discrete latent model of it. Our bisimulation bounds enable the application of formal methods for Markov decision processes. Finally, we show how one can use a policy obtained via state-of-the-art RL to efficiently train a variational autoencoder that yields a discrete latent model with provably approximately correct bisimulation guarantees. Additionally, we obtain a distilled version of the policy for the latent model.
Original languageEnglish
Publication statusUnpublished - 3 Jun 2022
EventBelgium-Netherlands workshop on Reinforcement Learning 2022 - Leiden University, Leiden, Netherlands
Duration: 3 Jun 20223 Jun 2022


ConferenceBelgium-Netherlands workshop on Reinforcement Learning 2022
Abbreviated titleBeNeRL 2022
Internet address


  • Reinforcement Learning


Dive into the research topics of 'Distillation of RL Policies with Formal Guarantees via Variational Abstraction of Markov Decision'. Together they form a unique fingerprint.

Cite this