An Efficient Hybrid Incremental Algorithm for Complex-Valued Feedforward Neural Networks

摘要

Complex back-propagation (CBP) algorithm based on gradient descent is one of the most popular optimization algorithms for training complex-valued feedforward neural networks (CVFNNs). However, it has the disadvantages of slow convergence and easy to fall into local minimum. Most of previous investigations on CVFNNs focused on the improvement of optimization algorithms while the structure was determined by the trial-and-error method, which is a time-consuming process. To solve this issue, an efficient complex-valued hybrid incremental algorithm (CV-HIA) is proposed in this paper for three-layer split CVFNNs. The size of the underlying CVFNN is adaptively determined by an incremental scheme. When the training error falls into local minimum, hidden neurons are added into the network one by one. Nonlinear parameters between neurons of the hidden and input layers are adjusted by the gradient descent method. For linear weights between neurons of the output and hidden layers, the least-square (LS) method is adopted. Experimental results on several benchmark datasets show that the proposed method not only achieves good performance, but also obtains a compact network structure.