Research On Solving Large-scale Optimization Problems Based On Fruit Fly Visual Evolution Neural Networ

As a kind of network structure with high parallel processing ability,neural networks have become a popular research branch in the fields of science and technology,but its wide application is restricted due to the difficulties of the network’s structure design and parameters,low optimization efficiency,and weak generalization ability.Interestingly,it has attracted more and more scholars’ attention to construct Fourier neural networks able to execute multi-dimensional data feature extraction with high execution efficiency and strong fitting ability.Also,it has been concerned to probe into visual evolution neural networks capable of handling optimization problems,relying upon the information processing structures and functional characteristics of the biological visual nervous systems and the idea of population evolution in swarm intelligence optimization.Flies as a special type of insect are extremely sensitive to the changes of moving objects in visual scenes,which provides a new biological inspiration for computer vision.Therefore,based on the classical Fourier neural networks and multilayer perceptron,the thesis firstly studies the multiinput Fourier neural network with strong noise suppression and feature extraction as well as high solution accuracy.Secondly,with the help of the visual neurophysiological theories and the inspiration of population evolution,visual evolutionary neural networks are here studied in order to solve large-scale global optimization and constraint optimization problems,in which their theoretical and experimental analyses are done,e.g.computational complexity and solution search performance.The acquired results play a certain role in promoting the development of evolutionary computation and computer vision,and provide new approaches for dealing with optimization problems as well.The main achievements acquired are summarized below:A.Aiming at the problem of feature extraction of multi-source time series,a gradient-based multiinput Fourier neural network is developed to execute data fitting and time series analysis.Thereafter,based on the difficulty of which the gradient descent method cannot acquire global optimal parameter settings,an improved sparrow search algorithm is developed to optimize its parameter settings and solve high dimensional function optimization problems.The theoretical analysis shows that the improved algorithm’s computational complexity is decided by its population size and the optimization problem’s dimension.Numerically comparative experiments have validated that,not only the acquired Fourier neural network can effectively extract the features of multi-attribute data with strong generalization ability,but also the improved algorithm has significant advantages in coping with high dimensional function optimization problems.B.In order to discuss the topic of large-scale global optimization(LSGO),a fly visual evolutionary neural network is developed to solve LSGO problems.It is composed of a fly visual neural network and a state matrix,in which any candidate solution is taken a state,while the values of the objective function are viewed as gray values used to construct gray images.Therein,based on the information processing mechanism of the fly visual neural system,the visual neural network is designed to output its activities named local and global learning rates;each state matrix is transferred into a new one in terms of a state transition strategy.Theoretical analysis shows that the computational complexity of the evolutionary neural network depends on the input size and the dimensionality of the optimization problem.Comparative experiments have verified that such a neural network is feasible and effective for LSGO problems.C.An improved fly visual evolutionary neural network is developed to solve single-objective constrained optimization problems.Herein,an improved fly visual neural network,which bases on the mechanisms of the fly’s visual persistence,visual attention and two-channel information processing,is designed to output a learning rate after each gray-scale image is inputted.Further,in the process of solution search,each state matrix is updated in terms of an improved strategy of population evolution.Theoretical analysis shows that the computational complexity of the evolutionary neural network is decided by the input size and the optimization problem’s dimension.Numerical experiments have showed that the visual neural network can well guide the current states to transfer toward potential regions,particularly it can adaptively generate penalty factors used to deal with the constraint conditions.At the same time,the visual evolutionary neural network can effectively handle constrained optimization problems with fast convergence and high accuracy.