Dynamic system modeling with extreme learning machines

Neural networks (NN) are used to find solutions to many engineering and science problems. The algorithms used for training these architectures are usually iterative algorithms. Derivative-based iterative algorithms are used to determine threshold and linkage weight values in feedforward neural networks. Derivative based iterative algorithms have a slow training period, which led to new searches. To overcome this slowness, the concept of extreme-learning machines (ELM) has made significant progress. ELM is a learning algorithm developed for single hidden layer feedforward networks. Although the ELM learning algorithm offers a great advantage in terms of training time, it cannot be said to have the same performance as generalization ability. This is why a new learning algorithm called Meta-ELM has been developed by combining traditional extreme-learning machines. In this study, dynamic system modeling performance of artificial neural networks trained with Meta-ELM was investigated. Training and test data sets were prepared for seven different dynamic systems selected from the literature to be used in this study. Meta-ELM system identification model was obtained for each dynamic system by using ELM training method. For each dynamic system, generalization achievements have been obtained with the test datasets of Meta-ELM models created with training phase. Modeling For the Meta-ELM, training and test datasets were repeated with different network parameters and the results were compared with the statistical results. Based on the results obtained, the performance changes of the proposed approach are shown. As a result, a general evaluation was made about the selection of the number of cells / nodes and number of ELMs in the group in order to construct the Meta-ELM and suggestions were made.