A Metacognitive Fully Complex Valued Functional Link Network for solving real valued classification problems

In this paper, a sequential learning based meta-cognitive fully complex valued functional link network (Mc-FCFLN) is developed for solving complex real world problems. Mc-FCFLN network consists of two components: a cognitive component and a meta-cognitive one. A fully complex-valued functional link network (FCFLN) is a cognitive component and the self-regulatory learning method is its meta-cognitive component. As the network does not possess hidden layers, the multi-variable polynomials are represented in the input layer for capturing the nonlinear relationship between the input and the output sample. Moreover, when the sample is presented to the Mc-FCFLN network, the meta-cognitive component decides what to learn, when to learn, and how to learn depending on the knowledge gained by the FCFLN network and the novel information present in the sample. The network can learn sample one after the other and thus the drawback existing with the batch learning strategy can be eliminated while orthogonal least square principle is used for selecting the best polynomial and the recursive least square update is used for tuning the network. Multi-category and binary datasets chosen from the UCI machine learning repository is used for the validation of the proposed classifier. Lastly, a performance comparison of the Mc-FCFLN applied for classification problems shows better classification ability when compared with the other existing classifiers in the literature.