Efficient Sequential Monte Carlo Using Interpolation

Abstract

A limitation common to all sequential Monte Carlo algorithms is the computational demand of accurately describing an arbitrary distribution, which may preclude real-time implementation for some systems. We propose using interpolation to construct a high accuracy approximation to the importance density. The surrogate density can then be efficiently evaluated in place of sampling the true importance density, allowing for the propagation of a large number of particles at reduced cost. Numerical examples are given demonstrating the utility of the approach.