The Use of Stochastic Optimization in Ground Motion Prediction

Abstract

In recent years stochastic optimization has become popular in modeling highly complex natural systems as opposed to traditional approaches, such as regression analysis. This paper tests the efficiency of regression analysis over mathematical optimization by evaluating the performance of stochastic (genetic algorithms, simulated annealing) and deterministic algorithms (nonlinear least squares, pattern search) in solving minimization problems. Qualitative and quantitative comparison of these solvers reveals that the implementation of genetic algorithms with initial population development, using Latin Hypercube sampling, outperformed other algorithms. Using the aforementioned solver, ground motion prediction equations are derived for the estimation of horizontal peak ground acceleration, velocity, and 5% damped spectral acceleration ordinates at discrete period estimators between 0.05 s and 2 s. Ground motion for stiff and soft soil sites is amplified by a constant factor with respect to rock site. These strong motion prediction equations are proposed for application in the range M4.5–M6.6 and for distances up to 150 km.