A Levenberg-Marquardt Algorithm Based Incremental Scheme for Complex-Valued Neural Networks

摘要

In this paper, an efficient learning algorithm is proposed for complex-valued neural networks. It can well tune the weights and bias by complex-valued Levenberg-Marquardt (LM) algorithm and optimize the network structure by an incremental mechanism. Firstly, compared with the first-order algorithm, such as complex-valued gradient descent algorithm, complex-valued LM algorithm overcomes the problem of slow convergence for training complex-valued neural networks. Secondly, for practical applications, the choice of optimization algorithms is over-emphasized, while the effect of the network structure on performance is ignored. Indeed, the complexity of the structure is crucial to the generalization capability. Therefore, in order to improve the performance of complex-valued neural networks, an incremental mechanism is adopted to determine the network structure during the training process. Experimental results show the effectiveness of the proposed algorithm.