Analysis and Clustering-Based Improvement of Particle Filter Optimization Algorithms

This study highlights how particle filter optimization (PFO) algorithms can explore objective functions and their robustness near optimums. Improvements of the general algorithm are also introduced to increase search efficiency. Population-based optimization algorithms reach outstanding performance by propagating not only one but many candidate solutions. One novel representative of these methods is the PFO concept, which was created as an analogue of the particle filter state estimation algorithm. The PFO algorithm results in a probability distribution of the sample elements, which can represent the shape of the objective function. In the literature, several variants of the PFO can be found, but its elements are not clearly fixed because of its novelty. In the present study, a method is introduced to gain information on the shape of the objective function by following the propagation of the particles along the iterations. The contributions of the paper: 1) comparative study is proposed examining the different variants of the algorithm, and some improvements are introduced (e.g., weight differentiation) to increase the efficiency of the general PFO algorithm; 2) propagation of the particles is investigated to explore the shape of the objective function; 3) clustering-based technique is proposed to get information about the local optimums (e.g., robustness). The results verify that the proposed method is applicable to find local optimums and evaluate their robustness, which is a promising prospect for robust optimization problems where often not the global, but a more stable local optimum gives the best solution.