Improvements on parsimonious extreme learning machine using recursive orthogonal least squares

Recently novel constructive and destructive parsimonious extreme learning machines (CP-ELM and DP-ELM) arose to cope with regression problems. With these foundations, several improvements on CP-ELM and DP-ELM are suggested. CP-ELM can be improved by replacing the Givens rotation with the Householder transformation, yielding the improved CP-ELM (ICP-ELM) which results in the acceleration of the training speed without hampering the generalization performance. Subsequently, a hybrid constructive–destructive ELM (CDP-ELM) is generated integrating elements from CP-ELM and DP-ELM. The goal is to combine the advantages of training speed and parsimony from CP-ELM and DP-ELM. Finally, experiments on regression data sets and a real-world system identification of robot arm example are done to test the feasibility and efficacy of these variants including ICP-ELM and CDP-ELM.