Quantum Evolutional Optimization And Deep Complex Neural Networks Learning

There are all kinds of complexity,nonlinearity,binding and other issues in the practical engineering applications.Now it is one of the main research direction and difficulty problem for finding a parallel intelligent optimization algorithm for large scale data analysis and many scholars pay extensive attention on this problem.At present,the heuristic intelligent optimization algorithm,which is based on biology intelligent or physical phenomenon,still exist many bottleneck problems in the practical applications.In this thesis,some improvement intelligent optimization methods based on quantum optimization and deep complex neural network has been proposed and applied to complex network community detection problem,polarization SAR image classification problem and image generation problem.Specifically,the major topics of the thesis are in the following.1.The community structure is an important feature of complex network,and we can find potential rules and phenomenon with mining community structure of network,that is help for understanding the relationship between the network structure and function.Research indicates that it has an important theoretical significance and wide application prospect.There are weak performance of space searching and too long consuming time in the existing algorithm for community structure of large-scale complex network.Therefore,quantum ant colony optimization algorithm(QACO-Net),combining ant colony algorithm with quantum computation,has been proposed effectively for large-scale community detection problem in the second chapter.In this algorithm,ant colony algorithm can make full use of the neighborhood information in complex network,which has good local searching ability.The quantum revolving gate is introduced for updating the pheromone of ant,which is help individuals to find a better solution.Meanwhile,it is effective to avoid falling into local optimum because of the uncertainty output of the quantum probability amplitude measurement.In addition,quantum ant colony optimization algorithm is a discrete algorithm which is suitable for this discrete problem of community detection.The experimental results shown that the proposed algorithm is effective for synthetic data and real world data.2.In the complex network community detection,modularity is the common index for measuring the community division.However,it has resolution limit problem with modularity optimization,which means a community smaller than a certain size cannot be detected with modularity optimization algorithm.In third chapter,a novel quantum discrete multiobjective particle swarm optimization algorithm(QDM-PSO)for community detection has been proposed,which is focus on resolution limit problem in large scale network.In this algorithm,quantum particle swarm optimization(QPSO)algorithm is used for updating the position and velocity of a particle.The algorithm has the advantage of fast convergence and less parameter to be optimized.The uncertainty of quantum mechanism makes the particle appearing in any location of the search space,so it has good global search ability.In addition,the multi-objective non-dominant selection strategy is adopted for overcoming the modularity resolution limit problem,so we can get the community divisions under different resolutions,then the decision makers will find the community division result with maximize value of modularity.The comparative experiments shown that QDM-PSO algorithm is not only effective for synthetic data and large-scale real world data,but also has high convergence speed.3.With the rapid development of remote sensing systems,the polarization of synthetic aperture radar(PolSAR)of complex scenario can get more and more rich PolSAR image data.How to interpret the massive PolSAR data quickly and efficiently is one of the challenging problems.As so for,deep learning has a good breakthrough in interpretation and understanding of remote sensing data,and most of them focus on real domain mainly,nevertheless,it is lacked of a complex domain model corresponding to the building blocks.In fourth chapter,an intelligent optimization method of deep complex convolution neural network is proposed,which makes full use of the representation ability of complex number and promotes robust for noise memory retrieval.This deep complex neural network contains complex weight initialization,complex convolution,complex maximum pooling,complex activation function,complex batch normalization,complex full connection,and complex output,etc.The experiments demonstrate that the deep complex convolution neural network has the feasibility and effectiveness of PolSAR image classification.4.Generative Adversarial Nets(GANs)is an intelligent optimization method,which is one of the main research topic in the field of artificial intelligence(AI)and the scholars of AI pay much attention to GANs.The neural network model consists of a generative model(G)and a discriminant model(D),in which the generative model can continuously capture the probability distribution of training samples and transform the input random noise into the generated samples.On the other hand,the discriminant model can discriminate the generate samples by observing the training sample and generate samples.With training,they can reach steady by competing with each other between training samples and generate samples.Now,GANs has been used to generate natural images,super resolution,text transform into image,image segmentation and classification,etc.In the fifth chapter,a deep complex domain convolution generation network is proposed,which is based on the depth of the convolution network model(DCGAN).The proposed model is the first time to expand the real domain model into complex domain convolution generation network,which is applied to image generation problem.Experimental results show that the deep complex generating network model has good result in image generation.