Develop adaptive tuning for robust online Markov-game learning

Develop adaptive tuning procedures for the robustness, exploration, and validation parameters of the optimistic dual robust Bellman framework without assuming that the relevant coverage or excitation constants are known in advance.

Background

The regret guarantees depend on parameters such as the exploiter-mix probability, validation radii, robustness level, and coverage or excitation constants. The main theory uses these quantities as specified inputs, and the paper notes that adaptation to unknown coverage or excitation constants is outside the analysis. A resolved treatment would provide guarantees while tuning these quantities online.