Dynamic Adaptive Search Strategy Based Incremental Extreme Learning Machine Based on

Extreme learning machine (ELM) is a promising method for the learning of single-hidden layer feedforward network (SLFN) which is attractive for its simplicity and high efficiency. However, during the rapid development of ELM algorithm, the determination of suitable network architecture is still a challenging work. To deal with this issue, this work develops a modified ELM algorithm based on a novel adaptive optimization method. Specifically, we use the growth structure strategy to design the network architecture. During the learning process of the proposed algorithm, the grey wolf optimization (GWO) technique is then introduced to seek the optimal parameters for hidden nodes instead of random selection. In addition, to improve the convergence speed, we further ameliorate the traditional GWO approach. Experiment results over some benchmark applications indicate that our AI-ELM algorithm can dramatically reduce the scale of network and obtains the better generalization performance than other classical ELM algorithms.