Prudent-Banker: No Extra Fees for Baseline Safety in Adversarial Bandits With and Without Delays
Researchers have introduced Prudent-Banker, a new algorithm designed for adversarial multi-armed bandits that maintains safety guarantees even with delayed feedback. This novel approach combines a delay-adapted Online Mirror Descent with a phased-aggression mechanism to ensure near-constant regret compared to a safe baseline policy. The algorithm's key innovation is a delay-calibrated restart threshold, which rigorously accounts for feedback distortions and reliably detects suboptimality. Prudent-Banker achieves optimal safety-robustness trade-offs, with theoretical guarantees and experimental validation showing its effectiveness in balancing safety and learning across various delay distributions. AI
IMPACT Introduces a novel algorithm for safe decision-making in complex bandit environments, potentially improving AI agents' reliability in real-world scenarios with uncertain feedback.