Investigation Into Application Of Hierarchical Genetic Algorithm To Training Of BP Network And To Blind Signals Separation

Encoding mechanism is a key part of genetic algorithm(GA). In general, a chromosome is coded as a string of genes. In this thesis, a new encoding mechanism is proposed by applying the thoughts of hierachical optimization. A chromosome is coded as a string of several segments, arranged in a hierachical structure, with each segments representing the different things.This kind of coding way, compared with the tranditional single layer encoding mechalism, would be able to represent the solution space more integratedly and accurately. Therefore it is more suitable for solving BP network training, TSP problem, Job scheduling etc.Currently, most of the algorithms for training BP network are splited into two discrete part, designing the structure of network and training the designed network. In the scenario posed in the first graph, a hierachical genetic algorithm is proposed for training BP neural network, by which the parameters—connection weights and thresholds(paramemeter genes) and the configuration of network—number of hidden layers and the number of neurons of each hidden layer(control genes) can be determined at the same time during training. In addition, we discuss a new solution to blind source separation (BSS). By encoding separation matrix as a two layers chromosome, and using the self-adapting characteristic of GA, mixed signals would be able to be separated without knowing numbes of sources, which is an indispensable prerequisite for most BSS algorithms. Tests are carried out respectively, and good performance of the new encoding mechalism of GA is demonstrated.