Symbolic Regression Based Extreme Learning Machine Models for System Identification

Abstract

Reproducible machine learning models with less number of parameters and fast optimization are preferred in embedded system design for the applications of artificial intelligence. Due to implementation advantages, symbolic regression with genetic programming has been used for modeling data. In addition, extreme learning machines have been designed with acceptable performances in virtue of random learning strategy. In this paper, symbolic regression featured extreme learning machine models are proposed for the system identification. The symbolic regression layer with mathematical operators and basis functions has been randomly constructed instead of genetic programming whereas the output weighting parameters are optimized via least-squares optimization as in extreme learning machines. Consequently; implementable, efficient and easy designed models are constructed for future applications. Comparative results of the proposed and literature models present that proposed models provided smaller mean-squared errors and minimum-descriptive length performances.