Research On Solving Large-scale Optimization Problems Based On Locust Visual Neural Network

Locust is an insect that is extremely sensitive to moving targets.Its visual system can perceive the moving behaviors of the targets in real time.This unique visual response characteristic provides a biologically theoretical foundation for computer vision.If the visual information processing mechanism of locusts are combined with the inspiration of population evolution,a new research branch,i.e.visual evolutionary neural network,will be born and become a potential research branch in intelligence computation.Therefore,with the aid of the locust's visual neurobiological theory and species evolution mechanisms,this work studies visual evolutionary neural networks for solving large-scale global optimization problems,and their computational complexity and comparative experiments are carried out.The work has a certain role in promoting the development of the integration of evolutionary computation and computer vision,and can also provide new solutions for engineering optimization problems.The main work and achievements are summarized as follows:A.Several art-of-the-state algorithms in the field of intelligence optimization are selected to execute comparative analysis,after which a multi-strategy hybrid evolutionary particle swarm optimization algorithm is proposed to cope with large-scale global optimization(LSGO)problems.The algorithm enhances its evolutionary ability through the strategies of symmetric learning and elite retention,and also the local Gaussian mutation strategy is introduced to strength the population diversity of the particle swarm.Theoretical analysis shows that the computational complexity of the algorithm depends on its population size and the dimensionality of the problem.Numerically comparative experiments have validated that the algorithm has the strong ability of evolution with satisfactory solution search performance,and thus is a potential optimizer.B.A locust visual evolutionary neural network algorithm is developed to solve LSGO problems.Precisely,relying upon the information processing mechanism of the locust visual neural system,an improved locust visual neural network is established to output its activity taken as a learning rate,in which the state matrix is composed of candidate solutions.Furthermore,a state transition strategy is designed to transfer the current states into new ones.Theoretical analysis shows that the computational complexity of the algorithm depends on the input size of the neural network and the dimensionality of the optimization problem.Comparative experiments verify that the algorithm is feasible and effective for LSGO problems.C.An improved locust visual evolutionary neural network algorithm is developed to enhance the performance of the above locust visual evolutionary neural network.More precisely,the idea of neural network layer division is adopted to divide the network's input layer into multiple blocks,each of which outputs an activity viewed as a learning rate.Hereafter,a state update strategy,which involves multiple inspirations of individual evolution,is designed to move the current states toward the optimal solution under the guidance of multiple learning rates.The algorithm's computational complexity is determined by the input size of the neural network and the dimensionality of the optimization problem.Experimental results show that the algorithm behaves well over the compared approaches with the aspects of efficiency,effect and universality.