A New Neural Network Based On Flower Pollination Algorithm

BP network is a error back propagation network.Because of its strong learning ability,flexible nonlinear modeling ability and large scale parallel computing power,it has been widely concerned by scholars,and has been widely used in various fields.But the problem of back propagation training is that the slow convergence and the chance to be trapped in local minima since it is a gradient based method especially when the samples are large and of high dimension.Flower pollination algorithm is a new metaheuristic algorithm for optimization.Because of it's advantages of simple structure,without gradient information,less parameters,easy to realize,has attracted more and more attention of scholars,and has been widely used in the fields of economic distribution,production scheduling.This paper presents the structure optimization of BP neural network based on the improved flower pollination algorithm,and the performance of the new network is verified by numerical experiments,through the analysis of the mechanism of BP algorithm and flower pollination algorithm.1.Flower pollination algorithm has some stagnation and thus a lower convergence rate under certain conditions,which can limit the search ability of the algorithm.For this reason and to improve the search efficiency,this paper proposes a self-adaptive mutation operator in the process of local pollination,self-adaptive adjustment of the switch probability,and hybridization of the flower pollination algorithm with the firefly algorithm,which leads to a new hybrid approach called self-adaptive flower pollination algorithm enhanced by the firefly algorithm.The proposed approach has been validated by benchmark functions and has been compared with other algorithms such as DE-FPA and PSO-FPA.Results indicate that the proposed hybrid algorithm has a higher rate of convergence and stability than other algorithms.2.The standard BP neural networks adjusts the weights and threshold value by the gradient descent method,which may be easily falling into a local optimum.Thus,to great extent,this may limit the overall optimization capacity.In this paper,FA-FPA is proposed and used to optimize the weights and threshold value of BP neural networks and the FA-FPA is integrated with the BP in two different forms: FA-FPA1-BP and FA-FPA2-BP.Finally,the performance of the standard BP network,FPA1-BP,FPA2-BP,FA-FPA1-BP,FA-FPA2-BP and GA-BP,PSO-BP is tested by using function approximation experiments and iris classification data set.The study shows that FA-FPA1-BP is superior to othernetworks in the function approximations and classification.