Sharp analysis of linear ensemble sampling
Researchers have published a new analysis of linear ensemble sampling (ES) in stochastic linear bandits, demonstrating its effectiveness with standard Gaussian perturbations. The study shows that ES can achieve a regret of \tilde O(d^{3/2}\sqrt n) with an ensemble size of m=\Theta(d\log n), matching the performance of Thompson sampling while maintaining comparable computational costs. The novel proof technique involves reducing the analysis to a time-uniform exceedance problem for independent Brownian motions, offering a new perspective on randomized exploration in linear bandits. AI