Reducing Calls to the Simulator in Simulation Based Inference (SBI)
Abstract
Simulation-Based Inference (SBI) deals with statistical inference in problems where the data are generated from a system that is described by a complex stochastic simulator. The challenge for inference in these problems is that the likelihood is intractable; SBI proceeds by using the simulator to sample from the likelihood. In many real world applications, simulator calls are expensive, limiting the associated sample size. Our goal in this work is to extend SBI to exploit two proposals for reducing simulator calls: to draw likelihood samples from a Neural Density Estimator (NDE) surrogate rather than from the stochastic simulator; and use of Support Points rather than simple random sampling to generate evaluation sites. We embed these methods in the Sequential Neural Posterior Estimator (SNPE) algorithm. Across a suite of test cases, we find that the NDE surrogate improves the quality of the inference; support points worked well in some examples, but not in others.