An autonomous incremental learning algorithm of Resource Allocating Network for online pattern recognition

In this paper, we propose a new autonomous incremental learning algorithm for radial basis function networks called Autonomous Learning algorithm for Resource Allocating Network (AL-RAN). The proposed AL-RAN can carried out the following operations autonomously: (1) data collection for initial learning, (2) data normalization, (3) allocation of RBFs, (4) setting and adjusting RBF widths, and (5) incremental learning. In this paper, we mainly improve the first four functions in the initial learning phase where a convergence criterion based on the class separability of collected data is adopted in order to reduce the computational costs. In AL-RAN, training data are first collected until the class separability is converged or the recognition accuracies for normalized and unnormalized data have a significant difference. Then, an initial structure of ALRAN is autonomously determined from the collected data, and AL-RAN is trained with them. After the initial learning, the incremental learning of AL-RAN is conducted whenever a new training data is given. In the experiments, we evaluate ALRAN using five benchmark datasets. The experimental results demonstrate that the above autonomous functions work well and the number of collected data in the proposed AL-RAN is significantly decreased without sacrificing the final recognition accuracy as compared with the previous version of AL-RAN.