COMPASS-Hedge: Learning Safely Without Knowing the World

A team of researchers has introduced COMPASS-Hedge, a novel algorithm that solves a fundamental trilemma in online learning. Online learning systems, which make sequential decisions based on feedback, have traditionally had to choose between excelling in unpredictable (adversarial) environments, efficient (stochastic) environments, or maintaining safety against a known baseline policy. Existing methods typically sacrifice optimal performance in one or two of these areas or require expert knowledge to tune parameters. COMPASS-Hedge is the first full-information method to unify all three goals without these compromises.

The algorithm's breakthrough lies in its novel integration of adaptive pseudo-regret scaling, phase-based aggression, and a comparator-aware mixing strategy. This allows it to automatically and simultaneously guarantee minimax-optimal performance against adversaries, instance-optimal (gap-dependent) performance in stochastic settings, and baseline safety with only logarithmic regret. Most importantly, it is parameter-free, meaning it requires no prior knowledge about whether the environment is adversarial or stochastic, nor the magnitude of performance gaps. This 'best-of-three-worlds' guarantee establishes that safety does not have to come at the cost of worst-case robustness or efficiency.

This work, published on arXiv, represents a significant theoretical advance with practical implications. By removing the need for environment-specific tuning, COMPASS-Hedge lowers the barrier to deploying robust and safe learning systems in real-world scenarios where conditions can change unpredictably. It provides a more reliable foundation for AI applications that must learn from sequential interactions, from dynamic pricing and recommendation systems to robotic control, without risking catastrophic failure by straying too far from a known safe policy.