Output-weighted sampling for multi-armed bandits with extreme payoffs

Abstract

We present a new type of acquisition function for online decision-making in multi-armed and contextual bandit problems with extreme payoffs. Specifically, we model the payoff function as a Gaussian process and formulate a novel type of upper confidence bound acquisition function that guides exploration towards the bandits that are deemed most relevant according to the variability of the observed rewards. This is achieved by computing a tractable likelihood ratio that quantifies the importance of the output relative to the inputs and essentially acts as an attention mechanism that promotes exploration of extreme rewards. Our formulation is supported by asymptotic zero-regret guarantees, and its performance is demonstrated across several synthetic benchmarks, as well as two realistic examples involving noisy sensor network data. Finally, we provide a JAX library for efficient bandit optimization using Gaussian processes.