A Novel Hybrid Algorithm of Particle Swarm Optimization and Evolution Strategies for Geophysical Non-linear Inverse Problems

Abstract

Population-based optimization algorithms are a class of stochastic global search methods, which probe the model space based on a set of nature-inspired rules in an iterative and random manner. Particle swarm optimization (PSO) is inspired by the social behavior of bird flocks and fish schools and is designed as a black-box optimization algorithm. In each iteration, a set of particles (i.e. potential solutions) simultaneously search the model space for each of which the cost function is calculated. Consequently, the computational cost of the search is directly related to the population size. Several empirical rules exist for the relationship between the model space dimension and the population size. But, since the model space dimension is problem-related, little discussion exists over the optimal population size for a successful convergence. However, compared to usual benchmark optimization problems, geophysical inversions have substantially higher dimensions, and as such large population sizes are mandatory for reaching meaningful solutions. Hence, the use of PSO becomes considerably infeasible for inverse problems in terms of the computation burden. Herein, a problem-oriented hybrid algorithm of PSO and evolution strategies, PSO/ES, is presented which integrates adaptive mutations into PSO with the goal of reducing the calculation time. As a result, instead of continuous search paths, particles follow a discrete scheme which allows them to search the model space in fewer numbers and more effectively. The algorithm is tested on a real 3D non-linear gravity inverse problem to estimate the thickness of the sedimentary cover in the South Caspian Basin. The problem is solved using both PSO and PSO/ES, where the results show that while PSO has prematurely converged due to insufficient population size, PSO/ES has been able to find a meaningful solution. The results agree well with the existing measurements in the study area.