| The neural network can predict to the complicated problem accurately ,but easy to affect by overtrain and trained slowly . The genetic algorithm is akind of the overall situation optimization algorithm of searching for, simplyand in common use, the characteristic such as being excellent and strong,however, one degree of function of fitting of it changes very big. The mainrequest for this method fits one degree of function that must disappear in theminimal error, this has left a lot of free space for concrete realization scheme.The ones that conbines the genetic algorithm and neural network together andcan raise the model on higher level are intelligible. "Intellectual systemintegration environment " is 863 project in the country undertaken inintellectual software of university of Jilin and knowledge engineeringlaboratory, the goal in research of the whole subject is to use and operate andsupport the research of the platform to Web intelligence systematically, bypiece, knowledge piece, space-time piece under data in the middle of visitingunder container, communication piece and data one, etc. make up betweenvisiting. This text has proposed that the improved genetic algorithmcombining with the method of excavating that LM optimizes the neuralnetwork, has developed the data in " intellectual system integrationenvironment " project and excavated the package on the basis of this method,we use this package to develop the tools of the maize to predict. The work ofthis text includes mainly:1. Have carried on detailed research on the genetic algorithm andartificial neural networkIncluding genetic algorithm prototype, the way of the genetic algorithmcode, go against the error and propagate the algorithm (BP algorithm) andneural network method based on algorithm of Levenberg-Marquardt(abbreviated as LM). Have discussed to the pluses and minuses of geneticalgorithm and artificial neural network.2. Combine the method of excavating that LM optimizes the neuralnetwork after putting forward the improved genetic algorithmHave consulted the combination method of genetic algorithm and neuralnetwork at present: First, optimize the neural initial right values of network;Second, adopt the neural network method to evolve (evolving neuralnetworks, is abbreviated as ENN), totally substitute BP to study with thegenetic algorithm, in order to avoid the defect of the descent method of thegradient. But there is a greater defect in these two kinds of combinationmethods: As to the former, genetic algorithm itself has early-maturingquestion of disappearing too, so this method can guarantee succeedingnetwork.It can fall into some little area extremely to train still. As to the latter,because of genetic algorithm own part ability weak characteristic , it evolvesneural network and needs one big initial right value area of range to make, itis unsatisfactory that the complexity arising from this rises and makes theneural network trained out suffused with the ability of melting, meanwhile,the population of the genetic algorithm calculates often makes its trainingexpenses much larger than the time expenses of BP algorithm.Because of above mentioned analysis, this text has proposed that theimproved genetic algorithm combining with LM optimizes the neuralnetwork method. Neural network algorithm is easy to fall into someadvantage most and restrains the slow shortcoming to feedforward type, thistext has introduced LM in non-linear least square method and optimizedalgorithms, this algorithm has strengthened some search ability, can improvethe precision of training of the network, shorten train time. Consideringgenetic algorithm in the course of searching for constantly may includeoptimum direction that solve searching for space to change, it is much biggercompared with simple neural network algorithm to search for the optimumprobability that solves of the overall situation, searching for the respect andnot so good as the neural network algorithm in the part , this text tries todivide two stages to use the genetic algorithm to improve the network to trainquality, carrying on through genetic algorithm to get one approximatesolution of the overall situation to adjust at first, regard this as initial value,adopt genetic algorithm and LM to optimize the neural network algorithmand run to train alternativly again, LM optimization neural network algorithmreaching to it can be through appointed precision or some vigorous stridesGenetic algorithm reaching LM optimization neural network algorithm canoperate through first intact genetic operator to switch over (choose theoperator, crossing operator, make a variation operator), it realizes to come, soeach other with training as one's own initial right value or result of the otherside, trains repeatedly and alternativly, until it reach the precision appointedor the most big step to count.Through carrying on the experiment in the typical data collected, we findthat it is better than other algorithms participating in being compared .Theimproved genetic algorithm combining with LM optimizing the neuralnetwork method (GALM algorithm), on the judging by accident rate, isslightly slow in HAAM algorithm in the restraining speed. The experimentindicates, though HAAM algorithm is quicker than GALM algorithm in therestraining speed, it is suffused with performance of melting to drop to someextent . And GALM algorithm, through using the genetic algorithm alone onthe initial stage of training to optimize the network algorithm and using withLM alternativly in middle period, can not only optimize the initial right valueof the network but also adjust the initial right value used in LM algorithmeach time effectively , meanwhile, LM algorithm strengthened some searchability of the genetic algorithm too;Making the network train extremelyavoid falling into, the network has the better one to suffused with the abilityof melting.3. Haveing developed the data and excavated the package because ofGALM algorithm, we use the output of the maize to predictOn the foundation of LM optimization of the neural network method ,theimproved genetic algorithm is put forward in this text, excavating thepackage after developing the data.And this package is the pretreatment of thedata, the result of data mining shows that three parts make up . Thepretreatment of the data is mainly partly to deal with the data list in thedatabase, the data lacked while expressing are mended completely, revise thewrong data;The data in the his-and-hers tables are excavated selectivelyaccording to the needs of user, if the attribute of the all right his-and-herstables is arranged , write down the conduct to simplify . The data excavateevery parameter that will show to users in the form of window body, for itsetting up according to the need , such as the precision of excavating , trainnumber of times ,etc. most largly. Excavate the result and show some , it willexcavate the result to show to users in the form of data list.We use this package to develop the output of the maize to predict tools,in temperate zone half moist LiShu county ,Jilin Province in recent ten yearsprecipitation , temperature and annual production of maize in ten years formonsoon climate of one year in order to the training data collection, withprecipitation, temperature, in order to the testing data collection .Precipitationand temperature , in order to input attributes in four years of 1993-1996years, the annual production of the maize, in order to output attribute, we usethis tool to excavate, to find that it is not people's mode correspondingrelation known to imply among the temperature , precipitation and annualproduction of maize in four years. After mining and analysing actually, wehave got the more ideal experimental result. Though it is difficult for themode knowledge received by this way to understand , we can regard it as "black box " to use , under having no guidance of any agricultural prioriknowledge , and carry on the prediction analysis of the output of the maize. |
PDF Full Text Request