Research And Application Of Transfer Learning

With the development of computers science and Internet,and the explosion of data,the technology of data mining and machine learning has achieved unprecedented development.Most machine learning algorithms rely on labeled data,but manual labeling is costly and labeled data is missing in many applications,so transfer learning came into being.We study the zero-shot learning in transfer learning,the application of transfer learning in unknown radar emitters recognition,and the application of reinforcement learning in radar jamming decision-making.Firstly,we study zero-shot learning based on linear compatibility function,improved the model by using a loss function that takes into account category differences,and refine the model by transductive local learning algorithm.The loss function of the existing zero-shot learning algorithms based on the linear compatibility function ignore the differences existing between different classes.Therefore,by using the differentiated loss function,the source domain model can better adapt to the classification task of the target domain.Then,transductive local learning algorithm is introduced,which uses the potential distribution information and structural characteristics of the target domain to reduce the problem of the projection domain shift,thereby improving the classification performance.The classification performance of the improved model has a certain degree of improvement.The local learning method can significantly improve the classification accuracy.Secondly,we study unknown radar emitter recognition algorithm based on transfer component analysis.We first analyze the problem of unknown radar emitter recognition and build the connection between known radar emitters and unknown radar emitters by expert knowledge.A novelty detection method based on a Gaussian mixture model is used to separate the known radar emitter data from the unknown radar emitter data in the test dataset.We use the transfer component analysis method to perform domain adaptation between known and unknown data and recognize type of radar emitter using a Bayesian Rule-based two-stage method.Finally,radar jamming decision-making algorithm based on Q learning is studied.We first analyze the key problem in radar jamming decision-making: the characteristics of radar modes,the influence of jamming on radar modes,and criterion of jamming effectiveness.Aiming at the characteristics of radar jamming decision-making problems,a corresponding reinforcement learning model was established.The target radar state corresponds to the state space,and the radar jamming type corresponds to the action space.State transition probability and reward function model are build based on different jamming effectiveness to different radar states.Then learn the jamming strategy autonomously using the model-free Q learning algorithm in the established radar jamming model.According to the requirement of rapid learning in radar confrontation,by introducing an adaptive learning rate,the convergence speed of the algorithm is improved.Experimental results show that the Q learning algorithm can learn the optimal jamming strategy in the environment,and the improved algorithm converges faster.