ICA model estimation using a mixed learning rule based on genetic algorithms and neural networks
Doru Constantin, Costel Bălcău, Doru Anastasiu Popescu · 2023
This paper presents a method of estimating the independent component analysis model with a training algorithm based on a mixed learning algorithm of genetic algorithms with a neural network algorithm. The mixed training algorithm is applied to optimize the objective function negentropy used to estimate the ICA model. The proposed estimation algorithm improves the training scheme based on genetic algorithms by using for crossover the most suitable chromosomes evaluated by the objective function with the parameters calculated calculated accordingly by a multilayer neural network algorithm. The results produced by the proposed algorithm were compared with the results of applying standard methods for estimating independent components such as FastICA variants based on the standard Newton method or the secant method. The experimental results were established in blind source separation applications by using unidimensional and bidimensional signals.