A genetic algorithm for blind source separation based on independent component analysis

This paper presents the implementation of genetic algorithm (GA) to a simple blind source separation(BSS) problem using independent component analysis(ICA). The process did not include pre-processing of mixture signals such as centering and whitening like most of ICA algorithms. The GA directly guesses the coefficients of the separating matrix given the mixture signals as inputs using maximization of kurtosis and minimization of mutual information as fitness function. Only one fitness function was defined to account the fitness for kurtosis and mutual information. Three set of simulations were performed. The first two simulations used the mixture of two and three synthetic signals, respectively. The third simulation used four audio signals. The results show that the proposed algorithm indeed separates the independent sources consisting of synthetic signals. The simulation consisting of four audio signals separates only three signals. It failed to extract one signal probably because the signal is almost a gaussian signal.