Kernel-based reinforcement Learning: A finite-time analysis

Omar D Domingues (1) , Pierre Ménard (2) , Matteo Pirotta (3) , Emilie Kaufmann (1, 4) , Michal Valko (5)

1 Scool - Scool
2 OVGU - Otto-von-Guericke-Universität Magdeburg = Otto-von-Guericke University [Magdeburg]
3 FAIR - Facebook AI Research [Paris]
4 CNRS - Centre National de la Recherche Scientifique
5 DeepMind [Paris]

We consider the exploration-exploitation dilemma in finite-horizon reinforcement learning problems whose state-action space is endowed with a metric. We introduce Kernel-UCBVI, a model-based optimistic algorithm that leverages the smoothness of the MDP and a non-parametric kernel estimator of the rewards and transitions to efficiently balance exploration and exploitation. Unlike existing approaches with regret guarantees, it does not use any kind of partitioning of the state-action space. For problems with $K$ episodes and horizon $H$, we provide a regret bound of O H 3 K max(1 2 , 2d 2d+1) , where $d$ is the covering dimension of the joint state-action space. This is the first regret bound for kernel-based RL using smoothing kernels, which requires very weak assumptions on the MDP and has been previously applied to a wide range of tasks. We empirically validate our approach in continuous MDPs with sparse rewards.

Domains: Machine Learning [stat.ML]

Omar D Domingues, Pierre Ménard, Matteo Pirotta, Emilie Kaufmann, Michal Valko. Kernel-based reinforcement Learning: A finite-time analysis. International Conference on Machine Learning, Jul 2021, Vienna / Virtual, Austria. ⟨hal-02541790v2⟩