Genetic evolution of radial basis function centers for pattern classification

This paper proposes to embed a genetic algorithm (GA) in the traditional learning algorithm of radial basis function (RBF) networks. Each function center of an RBF network is encoded as a binary string, and the concatenation of the strings forms a chromosome. In each generation cycle, the GA determines the center locations. Then the K-nearest neighbor heuristic and singular value decomposition are applied to find the function widths and output weights of each network, respectively. The performance of the proposed algorithm is evaluated on three problem sets. The results show that networks with centers found by the proposed algorithm achieve a lower mean squared error and a higher classification accuracy than networks with centers found by the K-means algorithm. The paper explains the findings by demonstrating that the best center locations may not necessary be located inside the input clusters.