An Incremental Extreme Learning Machine Prediction Method Based on Attenuated Regularization Term

Can Wang, Yuxiang Li, Weidong Zou, Yuanqing Xia

Abstract: As a powerful tool for regression prediction, Incremental Extreme Learning Machine (I-ELM) has good nonlinear approximation ability, but the original model has the problem that the uneven output weights distribution affects the generalization ability of the model. This paper proposes an Incremental Extreme Learning Machine method based on Attenuated Regularization Term (ARI-ELM). The proposed ARI-ELM adds attenuation regularization term in the iterative process of output weights, reduces the output weights of the hidden node in the early stage of the iteration and ensuring that the new nodes after multiple iterations are not affected by the large regularization coefficient. Therefore, the overall output weights of the network reach a relatively small and evenly distributed state, which would reduce the complexity of the model. This paper also proves that the model still has convergence performance after adding the attenuated regularization term. Simulation results on the benchmark data set demonstrate that our proposed approach has better generalization performance than other incremental extreme learning machine variants. In addition, this paper applies the algorithm to specific weight prediction scene of intelligent manufacturing dynamic scheduling, and also gets good results.