Drosophila Visual Neural Network Optimization Algorithm And Application

As a kind of network structure with high parallel processing ability,neural networks have played an important role in solving engineering problems.However,it is relatively sparse how to combine them with population evolution.Therefore,on the basis of exploring the information processing mechanism of the fly vision system,this dissert studies a new type of optimization approach after improving the existing fly visual neural networks,by combining such a type of neural network with either the gradient descent method or the inspiration of population evolution.The current research work not only helps for generating some interdisciplinary achievements between computer vision and intelligent optimization,but also provides parallel-processing schemes for optimization problems solving.The main work and achievements are summarized as follows:A.A fly visual neural network optimization algorithm is developed to solve bound-constrained optimization problems.Therein,the state matrix is constructed with the help of candidate solutions,and the objective function value of the state is regarded as a grayscale value.Hereafter,all the values of the states in the matrix form a gray image.On the other hand,based on the theory of fly visual neurophysiology,an improved fly visual neural network model is established to output its activity taken as a learning rate,while the gradient descent method is used to guide the state update.The theoretical analysis shows that the computational complexity of the algorithm is determined by the input size of the neural network and the dimension of the optimization problem to be solved.Numerically comparative experiments have validated that the algorithm can perform well over the compared approaches.B.Aiming at the problem of which the solution search performance of the above algorithm needs to be greatly improved,a fly visual evolutionary neural network algorithm,which bases on particle swarm optimization and the mechanism of fly visual information-processing,is designed to solve bound-constrained optimization problems.More precisely,an improved fly visual neural network with simple structure and few adjustable parameters is developed to generate its activity regarded as the learning rate of state transition.Further,a state updating strategy is designed to urge states to move toward the region which involves in the optimal solution,by means of the average position of the states.Then,under the guidance of the acquired learning rate,the state matrix is transferred according to the state update strategy.Numerically comparative experiments can draw a conclusion that the algorithm can acquire high-quality solutions with stable solution search.C.An improved fly visual evolutionary neural network algorithm is proposed to handle the problem of constraint function optimization.Precisely,an improved fly visual neural network,which can generate multiple output activities viewed as learning rates,is established by improving and simplifying the structure of the above fly visual neural network.A state update strategy with the characteristics of multi-strategy state update is designed to update the current states,depending on the strategies of glowworm position update and particle state update.Then,under the guidance of multiple learning rates,the states in the state matrix move toward the region which includes the optimal solution,based on the designed state update strategy.Numerically comparative experiments show that the algorithm has great advantages in the quality of the solution and the stability of solution search.