Cramér-Rao bounds on the performance of simulated annealing and genetic algorithms in EEG source localization

In this paper, we evaluate the performance of simulated annealing (SA) and the genetic algorithm (GA) when used for electroencephalographic (EEG) source localization. The performance is evaluated on the variance of the estimated localizations as a function of the optimization's initialization parameters and the signal-to-noise ratio (SNR). We use the concentrated likelihood function (CLF) as objective function and the Cramér-Rao bound (CRB) as a reference on the performance. The CRB sets the lower limit on the variance of our estimated values. Then, our simulations on realistic EEG data show that both SA and GA are highly sensitive to noise, but adjustments on their parameters for a fixed SNR value do not improve performance significantly. Our results also confirm that SA is more sensitive to noise and its performance may be affected by correlated sources.