Model Design And Performance Analysis Of Widely Linear Complex-Valued Extreme Learning Machine

In recent years,artificial neural networks have attracted more and more attention due to its great learning and data processing ability.Extreme learning machine(ELM)is an emerging learning model based on feedforward neural networks.The superiority of ELM in learning speed and generalization capability has been exhibited in literature over other conventional methods,such as error back-propagation algorithm(BP)which is slowly convergent and easy to be trapped into local minima.In order to effectively process the complex signals,complex-valued extreme learning ma-chine(CELM)has bee proposed by extending the real-valued extreme learning machine to complex domain.Indeed,the existing CELMs can not fully capture the second-order statistics information of complex signals.Thus,in most real-world applications,it's difficult to process the noncircular(or improper)signals.By introducing the conjugate information of complex in-put signals to CELM model,this thesis proposes two widely linear ELM models.The approx-imation abilities of the two proposed models are theoretically analyzed.Numerical samples show that the two proposed widely linear CELM models are more effective for both batch and online training than the original CELM.The first chapter mainly reviews some basic background knowledge about neural networks and extreme learning machine.Then,a brief introduction for original CELM is provided,in-cluding the network structure,learning algorithm and performance,and other relevant knowl-edge.By incorporating the conjugate information of the input and hidden nodes,respectively,two augmented CELM models are proposed in Chapter 2,where the approximation abilities of the proposed models are theoretically analyzed.In order to avoid the overfitting problem,the regularization method is introduced therein.To satisfy the need of online data processing,in Chapter 3,sequential training method of the two augmented CELM models are proposed.The above two proposed algorithms and the corresponding theoretical findings are verified with simulation examples,showing the effectiveness and superiority over the original CELM.