Optimizing the Topology and Learning Parameters of Hierarchical RBF Neural Networks Using Genetic Algorithms

摘要

This paper proposed a hierarchical topology of Radial Basis Function Neural Networks (HTRBFNs) combined with the using of evolutionary process by applying Genetic Algorithms (GAs). The GAs process used to optimize the topology of the HTRBFNs and its learning parameters. The number of sub-RBFNs in the hierarchical topology will be predetermined, and the GAs in its whole process will determine which of the input data variables group that directed to each sub-RBFN and it will optimize the sub-RBFN learning parameters (centers c, radii r and weights w). The proposed model implies that we can train exactly the system to optimize the topology and the learning parameters of each sub-RBFN depending on the entire model approximation error. GA-HTRBFNs produce automatically the most suitable topology which can apply to complex problems of function approximation. Therefore, the model goal is to find the most suitable topology of the proposed GA-HTRBFNs, the best optimization parameters and the number of RBF in each sub-RBFN in order to approximate a problem of input/output (I/O). The results show that the proposed GA-HTRBFNs model has the ability to produce the best topology of parallel RBFN structure using GAs, with the better value of approximation means square error.