2020/06/14 by S. Ashwin Renganathan, Renganathan, S. Ashwin, Jeffrey Larson +3
Computer Science · Decision Sciences · #Advanced Bandit Algorithms Research #Blind Source Separation Techniques #FOS: Computer and information sciences #FOS: Mathematics #Gaussian Processes and Bayesian Inference #Machine Learning (stat.ML) #Machine Learning and Algorithms #Optimization and Control (math.OC)
paper · pdf · doi:10.48550/arxiv.2006.08037
openalex publication_date 2020/06/14 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
We propose a novel Bayesian method to solve the maximization of a\ntime-dependent expensive-to-evaluate oracle. We are interested in the decision\nthat maximizes the oracle at a finite time horizon, when relatively few noisy\nevaluations can be performed before the horizon. Our recursive, two-step\nlookahead expected payoff (\r2LEY) acquisition function makes\nnonmyopic decisions at every stage by maximizing the estimated expected value\nof the oracle at the horizon. \r2LEY circumvents the evaluation of\nthe expensive multistep (more than two steps) lookahead acquisition function by\nrecursively optimizing a two-step lookahead acquisition function at every\nstage; unbiased estimators of this latter function and its gradient are\nutilized for efficient optimization. \r2LEY is shown to exhibit\nnatural exploration properties far from the time horizon, enabling accurate\nemulation of the oracle, which is exploited in the final decision made at the\nhorizon. To demonstrate the utility of \r2LEY, we compare it with\ntime-dependent extensions of popular myopic acquisition functions via both\nsynthetic and real-world datasets.\n