Sequential Monte Carlo Optimization Using Artificial State-Space Models

We introduce a method for sequential minimization of a certain class of (possibly non-convex) cost functions with respect to a high dimensional signal of interest. The proposed approach involves the transformation of the optimization problem into one of estimation in a discrete-time dynamical system. In particular, we describe a methodology for constructing an artificial state-space model which has the signal of interest as its unobserved dynamic state. The model is "adapted" to the cost function in the sense that the maximum a posteriori (MAP) estimate of the system state is also a global minimizer of the cost function. The advantage of the estimation framework is that we can draw from a pool of sequential Monte Carlo methods, for particle approximation of probability measures in dynamic systems, that enable the numerical computation of MAP estimates. We provide examples of how to apply the proposed methodology, including some illustrative simulation results.