Abstract
This survey gives an overview of Monte Carlo methodologies using surrogatemodels, for dealing with densities which are intractable, costly, and/or noisy.This type of problem can be found in numerous real-world scenarios, includingstochastic optimization and reinforcement learning, where each evaluation of adensity function may incur some computationally-expensive or even physical(real-world activity) cost, likely to give different results each time. Thesurrogate model does not incur this cost, but there are important trade-offsand considerations involved in the choice and design of such methodologies. Weclassify the different methodologies into three main classes and describespecific instances of algorithms under a unified notation. A modular schemewhich encompasses the considered methods is also presented. A range ofapplication scenarios is discussed, with special attention to thelikelihood-free setting and reinforcement learning. Several numericalcomparisons are also provided.