Performance comparison of optimization algorithms in matched field inversion

Optimization efficiencies and mechanisms of simulated annealing, genetic algorithm, differential evolution and downhill simplex differential evolution are compared and analyzed. Simulated annealing and genetic algorithm use a direct random process to search the parameter space for an optimal solution. They include the ability to avoid local minima, but as no gradient information is used, searches are relatively inefficient. Differential evolution searches the parameter space by using distance and azimuth between individuals of a population, although initial searches are effective, the search speed decreases quickly because differential information between the individuals gradually vanishes. Local downhill simplex and global differential evolution methods are developed separately, and combined to produce a hybrid downhill simplex differential evolution algorithm. The hybrid algorithm is sensitive to gradients of the object function and search of the parameter space is effective. These algorithms are applied to matched field inversion with synthetic data. Optimal parameter values, final values of object function and inversion time are presented and compared.