A practical mutation operator and its application to the Kalman filter

In this work we introduce a new mutation operator called the “selection follower (SF)” that exploits high eigenvalue-ratio and rotated-eigenvector functions. Unlike traditional mutation operators that scatter offspring with a fixed probabilistic distribution, the SF uses the shape of the population chosen by the selection operator as the probabilistic distribution in order to conform the offspring settlement to the fitness landscape. Experiments on test functions show that the SF is feasible both in search exploitation and exploration. Finally, the SF is applied to parameter estimation of a Kalman filter example that constitutes a 19-dimensional problem. Benchmarking with the expectation-maximization algorithm, the SF produces lower mean-square-estimates consistently. The robustness and feasibility of SF to practical problems are verified