An Improvement Sequential Minimal Optimization Algorithm Of Support Vector Machines

Support vector machines is one of the most important progress of machine learning in recent years.ItGot attention of many researchers both at home and abroad because of itsGoodClassification ability as soon as it appeared.It has been used in lots of fields plays a huge role.And all kinds of support vector machine algorithm also have been the scholar's research focus,especially the sequential minimal optimization algorithm.Sequential minimal optimization algorithm has beautiful expression of quadratic programming which avoids excessive space requirements,and makes the realization of support vector machine be simple and efficient.But the efficiency problem of sequential minimal optimization algorithm is also the focus ofCurrent research.In this paper,we analyzed the solving process of sequential minimal optimization algorithm and support vector machine by a large number of experiments with the help of theChange of the objective function value and the interval value on the basis of the traditional sequential minimal optimization algorithm.And also improved the stopCondition of the process of algorithm according to the analysis results.During the improvement process we smoothed the objective function value and interval valueChangeCurve.Then statistical data to measure both the effect of the improved sequential minimal optimization algorithm.And adopt the method ofCross validation verify the results of the improved algorithm.The research work of this paper mainly includes:(1)Deduced expressions of according to the traditional sequential minimal optimization algorithm the objective function value and interval value whichCan respectively output target function value and the data of interval valueChanging with the number of iterations.Then observe the quantity of eachCase during the process.(2)Smoothed theCurve of theChange of objective function value and intervals during the experimental process,and showed theChange process with more visual way.Then found the rule of objective function value and interval valueChanging with the number of iterations.Abundant experimental results indicated that theChange was in hinge function form.AfterCertain number of iterations,objective function decreases very little in a long period of time rest.Sometimes even small increasing/decreasing fluctuation appears.(3)Improved the process of the traditional sequential minimal optimization algorithm.We did the improvement in both objective value and interval value.Find a stop standard whichCan terminate training in advance to avoid inefficient training and have little effects on the accuracy of the training at the same time.Then finished theCode.(4)Did the experiments respectively for the two improved sequential minimal optimization algorithm experiment,and statistics data.ThenCompared the experimental results of the traditional algorithm from the aspects of training efficiency and test accuracy of improved algorithm.At the same time,more authoritative Cross validation method is used for training efficiency and model prediction ability.Analyzed the training effect of the improved algorithm synthetically.Through a lot of experiments,the new improved minimum sequence optimization algorithm for this paper proposed is superior in the training efficiency and model prediction ability.And margin value assisted sequential minimal optimization improvement is more significant on improving the training efficiency.