Toward Optimal Regret in Adversarial MDPs with Stochastic Hard Constraints

Name: Toward Optimal Regret in Adversarial MDPs with Stochastic Hard Constraints
Uploaded: 2026-10-15T00:46:54.907475+00:00
Duration: 437 s
Description: MA-OPS is an algorithm for learning policies in adversarial Markov Decision Processes (MDPs) with hard constraints. It combines searching for a policy with a large safety margin using an optimistic search with evaluating selected policies pessimistically. This approach yields a regret bound of Oe(√T /ρ + 1/(dρ)), ensuring safety while minimizing loss.

MA-OPS is an algorithm for learning policies in adversarial Markov Decision Processes (MDPs) with hard constraints. It combines searching for a policy with a large safety margin using an optimistic search with evaluating selected policies pessimistically. This approach yields a regret bound of Oe(√T /ρ + 1/(dρ)), ensuring safety while minimizing loss.

Open the full episode