New Algorithm Improves Linear Bandit Exploration Efficiency

In the realm of stochastic linear bandits, algorithms like Upper Confidence Bound (UCB) offer strong theoretical guarantees but can be computationally intensive. Conversely, Thompson Sampling (TS) is efficient but often harder to analyze due to its non-optimistic nature. This research bridges this gap by proposing a new approach. The paper introduces Absolute Thompson Sampling (ATS), a variant of TS that replaces signed exploration noise with its absolute value. This simple change ensures optimism in expectation, allowing for a UCB-style regret analysis while retaining TS's computational benefits. ATS achieves competitive regret bounds, matching existing TS results. Furthermore, the authors present Ensemble Absolute Thompson Sampling (EATS), which aggregates multiple absolute perturbations. EATS demonstrates strong performance with moderate ensemble sizes and theoretically converges to UCB behavior as the ensemble grows, offering a practical and theoretically sound method for balancing exploration and exploitation.