A Jointly Efficient and Near-Optimal Algorithm for Heteroskedastic Generalized Linear Bandits with Adversarial Corruptions
The HCW-GLB-OMD algorithm addresses heteroskedastic generalized linear bandits facing adversarial corruptions. It combines online mirror descent with Hessian-based confidence weights to achieve instance-wise minimax optimality. This provides a robust framework for sequential decision-making when rewards follow complex models subject to adaptive manipulations.