Modelling of suspended matter by hybrid RBF-IGNG network

The aim of this paper is to look for efficient algorithms allowing to deal with taking into account the modelization and the cartography of the sea components. Through the analysis carried on the family of Growing Neural Gas with an evolutive (scalable) architecture and with non supervised competitive learning, some modifications and improvements (modified IGNG) have been brought in order to associate them with the neural network of modelization; this modification is essentially focused on the automation of parameters. So as to improve the results, we propose in our technique to widen the sphere of influence of the RBF by multiplying the Gaussian widths by a factor which is automatically sought for so as to minimize the error of modelization on the learning basis. A study on the type of Gaussian widths of the REF and their shapes has been carried. The developed methodology, the implemented procedures and the proposed networks all yielded satisfying results.