Uncertainty of parameters in structural identification by genetic algorithm

Abstract

Uncertainty of model parameters identified by Genetic Algorithm (GA) is addressed in this study. A framework to estimate the uncertainty of parameters is formulated based on the specified threshold value of fitness. In order to interpret the identified results and to estimate the uncertainty of parameters, Monte-Carlo Simulation (MCS) can be adopted. Nevertheless, performing many repetitions of GA in MCS is computationally expensive, especially for systems with many unknown parameters. For this purpose, a GA based adaptive search algorithm that employs adaptive search space zooming strategy is proposed, and later adopted to explore the optimal neighbourhood of the fitness evaluation function. An ensemble of solutions with equal or higher fitness than threshold is used to evaluate the statistics of identified parameters under normal PDF, or Skew-Normal PDF whichever the data is fitted to. Confidence intervals of parameters are calculated from these distributional parameters. Numerical results show good interpretation of identified results with considerable saving in computational time in comparison with the brute-force approach of repeating GA runs.