Identification of Blackwell Policies for Deterministic MDPs

Victor Boone (1) , Bruno Gaujal (1)

1 LIG - Laboratoire d'Informatique de Grenoble

We consider the problem of the identification of Blackwell optimal policies for deterministic finite Markov Decision Processes (d-MDPs). Specifically, we are interested in algorithms that learn reward distributions by querying samples over time, that stop almost surely and return a Blackwell optimal policy with high probability. We provide a characterization of the class of MDPs over which such algorithms exist together with an algorithm identifying Blackwell optimal policies with arbitrarly high probability.

Mots clés: Reinforcement Learning; Markov Decision Processes; Blackwell Optimality

Domaines: Recherche opérationnelle [math.OC]; Combinatoire [math.CO]

Victor Boone, Bruno Gaujal. Identification of Blackwell Policies for Deterministic MDPs. 23ème congrès annuel de la Société Française de Recherche Opérationnelle et d'Aide à la Décision, INSA Lyon, Feb 2022, Villeurbanne - Lyon, France. ⟨hal-03595301⟩