Particle Filter Optimization: A Bayesian Approach for Global Stochastic Optimization

This paper introduces a novel global optimization algorithm called Particle Filter Optimization (PFO), designed for a class of stochastic problems. PFO leverages the Bayesian inference framework of Particle Filters (PF) by integrating the optimization problem into the PF estimation process. In this context, the objective function replaces the measurement, and a customized transitional prior is developed to function as state dynamics. This dynamic replaces classic acquisition function and grants the PF a local optimization capability, facilitating its transformation towards global optimization. In PFO, the particles serve as agents in the optimization problem. Given the noisy nature of measured outputs, the Unscented Transform (UT) is utilized to estimate the true mean, thereby reducing the impact of erroneous information on particle transitions and weight updates. The algorithm is designed to minimize the introduction of unnecessary parameters and adheres to theoretically validated PF procedures, resulting in a robust heuristic algorithm supported by rigorous theoretical foundations.