MultiLearner Based Recursive Supervised Training

In supervised learning, most single solution neural networks such as constructive backpropagation give good results when used with some datasets but not with others. Others such as probabilistic neural networks (PNN) fit a curve to perfection but need to be manually tuned in the case of noisy data. Recursive percentage based hybrid pattern training (RPHP) overcomes this problem by recursively training subsets of the data, thereby using several neural networks. MultiLearner based recursive training (MLRT) is an extension of this approach, where a combination of existing and new learners are used and subsets are trained using the weak learner which is best suited for this subset. We observed that empirically, MLRT performs considerably well as compared to RPHP and other systems on benchmark data with 11% improvement in accuracy on the spam dataset and comparable performances on the vowel and the two-spiral problems