Finding Optimal Parameters for Neural Gas Networks Using Evolutionary Algorithms

The parameter values used for the Growing Neural Gas (GNG) algorithm are generally determined empirically. This requires long calculation times and may lead to values which are not optimized for the data set they are being used with. The present work proposes the use of Evolutionary Algorithms to optimize these parameter values. During the optimization process, GNG networks are created with the parameter values stored in individuals from the Evolutionary Algorithm, and trained with object features extracted from images of the ETH-80 database. An individual’s fitness is calculated according to three different functions which assess the performance of the trained GNG networks. A feature-based object recognition and categorization model defined by a taxonomic hierarchy of self-organized GNG networks is trained and tested using the parameter values obtained from the optimizer, and with the same data sets employed during the optimization process. The categorization and recognition rates of the obtained models are compared to the ones achieved with the empirically set parameter values. The results suggest that the optimization of parameter values significantly increases the model’s performance when using the constrained initial values for the individuals in the first parent population of the Evolutionary Algorithm. Using unconstrained initial values seems to lead to parameter values that, when used in the GNG algorithm, create unstable networks, which are incapable of learning. This is likely caused by the existence of local optima throughout the parameter search space.