The ELM learning algorithm with tunable activation functions and its application

Based on the problem dependency of activation functions with Extreme Learning Machine (ELM) learning algorithm, the ELM learning algorithm with tunable activation functions is proposed in this paper. The presented algorithm determines its activation functions dynamically with differential evolution algorithm based on the input training data of problem. Compared with ELM and E-ELM learning algorithms for benchmark problems of function approximation and pattern classification, the simulation results show that the proposed algorithm can provide better generalization performance and robustness with the same network size and compact network structure.