A variable-structure sequential ELM algorithm based on the characteristics of time-varying system
Licheng Cui, Huawei Zhai, Zengtang Qu, Benchao Wang · 2017
Extreme learning machine (ELM) is an efficient algorithm, the number of hidden neurons is the key point, it has a great effects on the final results and the performances of the algorithm, but its number of hidden neurons is key and difficult point to determine, if the number is too large, it will reduce the accuracy, otherwise, it will lead to be over fitting. So, by analyzing the characteristics of the training data, an improved ELM algorithm is proposed, CI-ELM, it could adjust the number of hidden neurons by the sensitivity analysis dynamically, so as to adjust the structure of the network. Contrast tests based on the time series data show CI-ELM is superior in training speed and error, and it is acceptable.