Weighted Linear Bandits for Non-Stationary Environments
This work addresses the challenge of adapting to changing environments in online decision-making for applications like recommendation systems, though it is incremental as it builds on existing linear bandit frameworks.
The paper tackles the problem of non-stationary linear bandits, where the unknown regression parameter varies over time, by proposing D-LinUCB, an optimistic algorithm based on discounted linear regression with exponential weights to forget past data. It provides theoretical guarantees, including an optimal dynamic regret bound of order d^{2/3} B_T^{1/3} T^{2/3}, and demonstrates empirical performance in simulations.
We consider a stochastic linear bandit model in which the available actions correspond to arbitrary context vectors whose associated rewards follow a non-stationary linear regression model. In this setting, the unknown regression parameter is allowed to vary in time. To address this problem, we propose D-LinUCB, a novel optimistic algorithm based on discounted linear regression, where exponential weights are used to smoothly forget the past. This involves studying the deviations of the sequential weighted least-squares estimator under generic assumptions. As a by-product, we obtain novel deviation results that can be used beyond non-stationary environments. We provide theoretical guarantees on the behavior of D-LinUCB in both slowly-varying and abruptly-changing environments. We obtain an upper bound on the dynamic regret that is of order d^{2/3} B\_T^{1/3}T^{2/3}, where B\_T is a measure of non-stationarity (d and T being, respectively, dimension and horizon). This rate is known to be optimal. We also illustrate the empirical performance of D-LinUCB and compare it with recently proposed alternatives in simulated environments.