| Traditional machine learning techniques make a basic assumption that the trainingand test data should be under the same distributions, if the distribution of experimentaldata has changed, we need to re-label new samples as a training set to establish a newstatistical modal, however new data are expensive to label. Transfer learning can beused to reduce requirements of new marked data. We can improve the precision ofmodel by transferring auxiliary marked data of different fields to target area. In thispaper we investigate the transfer learning theory to improve the performance ofrecognition on radar radiant fields.Firstly, RBF neural network algorithm is implemented for radar radiant recognition.When very few samples in training set, the precision of trained classifier for radarradiant recognition is low. For this problem, we carefully researched transfer learningtheory, and design a novel RBF neural network classifier based on transfer componentanalysis, the system reduces the identification error by transferring data in differentfields to target area to support the training of classification. Secondly, we investigatedthe ensemble learning theory and AdaBoost algorithm in the framework of the transferlearning, and then design the TCABoosting algorithm as an improved AdaBoostalgorithm, experimental results show that the reformed algorithm meliorates theaccuracy of the system. Finally, the relationship between performance of classificationand distribution adjusted function in Boosting algorithm is explored, and then weemploy Correction function that dynamically adjust the distribution in TCABoostingalgorithm to improve the stability of the system. |
