Florent Delgrange, Ann Nowe, Guillermo A. Pérez
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.
Delgrange, F, Nowe, A & Pérez, GA 2022, 'Distillation of RL Policies with Formal Guarantees via Variational Abstraction of Markov Decision Processes: Encore Abstract', BNAIC/BeNeLearn 2022, Lamot Mechelen, Belgium, 7/11/22 - 9/11/22.
Delgrange, F., Nowe, A., & Pérez, G. A. (2022). Distillation of RL Policies with Formal Guarantees via Variational Abstraction of Markov Decision Processes: Encore Abstract. Abstract from BNAIC/BeNeLearn 2022, Lamot Mechelen, Belgium.
@conference{036ca921ab124e1c839f2563b7851efb,
title = "Distillation of RL Policies with Formal Guarantees via Variational Abstraction of Markov Decision Processes: Encore Abstract",
abstract = "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.",
keywords = "Reinforcement Learning, Formal Methods, Representation Learning",
author = "Florent Delgrange and Ann Nowe and P{\'e}rez, {Guillermo A.}",
year = "2022",
month = nov,
day = "7",
language = "English",
note = "BNAIC/BeNeLearn 2022 : Joint International Scientific Conferences on AI and Machine Learning, BNAIC/BeNeLearn 2022 ; Conference date: 07-11-2022 Through 09-11-2022",
url = "https://bnaic2022.uantwerpen.be/",
}