EFFICIENT SAMPLING BUDGET ALLOCATION FOR ONLINE DECISION MAKING

Abstract

Often, executing simulation experiments is time-consuming, which hinders its application in real-time decisions. We firstly consider a problem where observations are censored. We formulate it as ranking and selection with censored observations and propose a convergent sampling algorithm. Then, we allow performance to vary with scenarios, where outcomes can be predicted via a logistic model. The goal is to identify feasible alternatives conditional on scenarios. An information gradient policy is developed to allocate offline simulation budget. Lastly, we consider a short time after observing each online scenario and propose a unified offline and online learning paradigm to select the best conditional on online scenarios. We model the mean performance as a function of scenarios and learn a predictive model based on simulation observations. Then, we develop a sequential sampling method to allocate online simulation budget. The superior performance of proposed approaches is demonstrated on synthetic functions and real-world applications.