Improved GA-BP Algorithm In The Fitting Of GPS Eevation

The artificial neural network is an intelligent numerical mapping modelof computation which is formed according to the information transfer betweenneurons and the mode of processing,to establish the relationship between the inputand output values through the connection weights between neurons and thresholds.There is no limit on the dimension of the input and output values. The artificial neuralnetwork has a wide range performance of data models and is widely used in thenumerical fitting and forecasting. BP neural network is an artificial neural network, byusing the gradient descent method of the error function to the output value, search forthe optimal value of connection weights and thresholds. BP network has strong abilityto Pan and value mapping, but there are many inadequacies, mainly in the uncertaintyof the network structure, it is difficult to determine the most appropriate hidden layersand nodes number of the network for the actual problem, which will directly affect theoutput accuracy. BP network training process use gradient descent method, if the stepsize is too large, the network training can not be convergence; the use of a smallerlearning rate, the network is easy to fall into a very small value of error function, cannot get the optimal weights and thresholds of the network. To solve this problem inthis article use the improved genetic algorithm to optimize the BP network.The genetic algorithm is an intelligent search algorithm through the simulation ofbiological evolution, by using genetic operators to adapt to the optimal direction ofsearch for the value of fitness function, is a strong global optimization searchalgorithm. The simple GA algorithm through the evaluation of the individual fitnessin population to change the direction of the evolution of the population, after a certainnumber of evolutions, takes the individual which has largest value of fitness functionto adapt to the optimal solution value of the problem. The hybridization and mutationrate in Genetic operators is directly related to the fate of the individual, determines theability and direction to produce new populations. Under normal circumstances, Pc andPm values can only be obtained through similar problems spreadsheet or using the experience of the Pc, Pm value. From the point of biological evolution in this article,build adaptive regulation system of genetic algorithm. Individual fitness value is moreconcentrated, the probability to generate new individuals should be larger, otherwisethe probability should be more scattered, so that it can help solve the globaloptimization.Simulated Annealing algorithm for the simulation of liquid metal crystallizationprocess obtains the minimum energy value by sequential cooling. Metropolisacceptance criteria in SA algorithm applied to the genetic algorithm, the individual ofpoor fitness is accepted a certain degree of probability, so that we can ensure thediversity of individuals in a population and inhibit premature convergence of GAalgorithm, at the same time as the temperature decreases, the probability of accept alower fitness individual is low, eventually converge to the optimal individual.Beginning of the algorithm, the value of individuals is divergence, it is in globalsearch favor, the latter part of SA algorithm, the value of individuals is focused, it canhelp speed up local search. This improvement is more conducive to global and localsearch capabilities of the GA algorithm.Encoding of the network weights and thresholds, take the BP network valuedomain into the GA algorithm mode domain, use the genetic algorithm to search theglobal optimal value of the BP network output errors, change the gradient descentsearch of simple BP neural network to avoid the error of BP network search intolocalization. This article will use this theory to the fitting of GPS Elevation, the choiceof the three-dimensional coordinates of GPS stations as an abnormal elevationdirectly impact factor, use the BP neural network to establish a mathematical model ofGPS Height Transformation. Use the encoding of the network weights and thresholdsto establish contact BP and the GA algorithm. Combine Genetic algorithm andgradient descent method to search the network weights and thresholds of BP neuralnetwork output error minimum. Save the optimal weights and thresholds of BP neuralnetwork training. Change the input data corresponding to the predicted values.