A Meta-cognitive Fully Complex Valued Functional Link predictor Network for solving real valued prediction problems

摘要

In this paper, a Meta-cognitive Fully Complex Valued Functional Link predictor Network (Mc-FCFLNP) is developed for solving the complex practical problems. Mc-FCFLNP network contains two components, first, a cognitive and next a meta-cognitive component. A Fully Complex-valued Functional Link network (FCFLNP) acts as a cognitive component and its self directed learning mechanism acts as meta-cognitive component. As the network does not possess hidden layers, the multi-variable polynomials are used in the input layer for representing the non-linear relationship between the input and the output. When the sample is sent to the Mc-FCFLNP network for training, the meta-cognitive component chooses what-to-learn, when-to-learn, and how-to-learn depending on the knowledge attained by the FCFLNP network and the novelty of the sample. The network utilises the sequential learning methodology for eliminating the limitations existing with the batch learning strategy. The recursive least square (RLS) update is used for tuning the output weight of the network and the Orthogonal Least Square (OLS) principle is used for the selection of the best polynomial. A set of bench mark prediction problems are used for validating the proposed network. Performance comparison of the Mc-FCFLNP clearly shows a better prediction ability when compared with the other existing networks in the literature.