Performance of different metaheuristics in EEG source localization compared to the Cramér–Rao bound

The problem of electroencephalographic (EEG) source localization involves an optimization process that can be solved through metaheuristics. In this paper, we evaluate the performance in localizing EEG dipole sources using simulated annealing (SA), genetic algorithm (GA), particle swarm optimization (PSO), and differential evolution (DE). The evaluation is performed in terms of the metaheuristics' operational parameters and the signal-to-noise ratio (SNR). Here, the objective function in the optimization problem is the concentrated likelihood function (CLF), while the Cramér–Rao bound is proposed as benchmark. Under these conditions, the metaheuristics' best performances are achieved when the variance of their corresponding dipole source estimates is close to the CRB. Therefore, we exhaustively evaluate the variances of the source estimates obtained through each metaheuristic for the case of realistically simulated EEG data. Our results showed that no significant variations on the performance were introduced by changes in the metaheuristics' operational parameters, specially in one-dipole estimation. For two-dipole estimation, the performance of GA and DE was better than in other metaheuristics, but DE required a fine adjustment of its parameters in order to work. In all cases, the performance decayed as the SNR decreased, while SA and PSO seemed to be very sensitive to the correlation between the sources. Overall, GA was the most attractive technique in terms of performance and computational cost.