Partial safety with optimal learning beyond stationary environments

Determine whether there exists an online learning algorithm that achieves partial safety while maintaining optimal learning performance across a broader class of environments beyond stationary settings, extending robustness and learning guarantees beyond the stationary case analyzed in this paper.

Background

The paper demonstrates that partial safety is incompatible with no-external-regret but feasible with no-weak-external-regret, and provides algorithms (EEP/ESEP) that achieve strong performance in stationary environments while safeguarding against adaptive opponents.

The authors highlight the need to extend these guarantees beyond the stationary setting. Identifying algorithms that remain partially safe while still learning optimally in broader, potentially non-stationary or strategic environments would deepen understanding of the limits and possibilities of robust online learning under adverse selection.