Independent component analysis based on genetic algorithms

FastICA and Infomax are the most popular algorithms for calculating independent components. These two optimization process usually lead to unstable results. To overcome this drawback, a genetic algorithm for independent component analysis has been developed with enhancement of the independence of the resulting components. By modifying the FastICA to start from given initial point and adopting a new feasible fitness function, the original target of obtaining the maximum mutual independence is achieved. The proposed method is evaluated and tested on a numerical simulative data set from the measures of the normalized mutual information, negentropy and kurtosis, together with the accuracy of the estimated components and mixing vectors. Experimental results on simulated data demonstrate that compared to FastICA and Infomax, the proposed algorithm can give more accurate results together with stronger independence.