Reinforcement One-shot Active Learning Algorithms For Incomplete Supervision Data And Its Application

With the advent of the era of big data,how to effectively label massive data has become a hot issue.In practical applications,labeled data are scarce whereas unlabeled data are abundant.Due to the involvement of human experts,the cost of tag acquisition is usually high,and the cost of obtaining all monitoring information is large.As a new machine learning method,active learning can reduce the workload of human annotation in the machine learning system by querying the most valuable samples.It can reduce the cost of labeling without sacrificing the performance of the model,and has broad application prospects.Although the traditional active learning method has achieved gratifying results in the classification of less categories such as binary classification,the application research of active learning in the background of big data problems has been faced with many challenges.First,because many active learning query strategies need to perform matrix inversion and other operations,the amount of calculation increases exponentially with the increase of the scale of the problem,and it is difficult to apply in multi-category data.Second,the traditional active learning strategies mostly rely on inspiration.The method or theoretical metric,because the sample distribution in online-learning is unknown,the data distribution of potential learning problems is likely to change with time,and the traditional active learning method is often incompetent.Aiming at solving the difficulty of applying traditional active learning in multi-classification problems,we design an active learning model based on subsampling.Through experiments on the Binary Alphadigits and OMNIGLOT datasets,we have demonstrated the feasibility of the subsampling method for active learning to solve multi-category classification problems.At the same time,through the analysis of the experimental results,we find that although the subsampling method can break the limitation that the existing traditional active learning method cannot deal with the classification problem of large-scale data,there are still problems such as low classification accuracy and too many queries.In order to solve the problem that the traditional active learning can not actually adjust the sample query strategy online according to the problem,we designed a small sample based enhanced active learning model(ROAL).The experiment shows that ROAL can speed up the training by introducing the cross entropy term into the loss function,effectively avoiding the inefficient phenomenon and gradient disappearance at the beginning of training,and achieving the purpose of saving time and computing resources.Secondly,we designed the Binary Alphadigits data set and OMNIGLOT data set as the problem structure of online learning,and verified the effectiveness of ROAL model in online learning problems.In addition,by designing the sequence of samples,it is verified that ROAL model has learned an effective action strategy and can decide whether to query the label of this sample or predict the label of this sample according to the uncertainty.Finally,the effects of the setting of reward value function on the performance of the reinforcement active learning model and the change of the performance of the reinforcement active learning model when the number of categories changes are discussed through experiments.Based on the problem background of battlefield aircraft identification,we studied the application of active learning in aircraft recognition based on situational data for the case where the distribution of test samples and training samples is quite different(i.e.the recognition of new aircraft models that have not been seen before).Firstly,a feature extraction method based on motion performance was designed for the real aircraft track data of one month in a certain airspace,and the effectiveness of feature extraction was verified through the comparative experiment on the traditional supervised learning model.Secondly,in view of the limitation of traditional active learning methods that are not competent for multi-classification problem,we study the application of subsampling based active learning model in aircraft recognition.Finally,we use the small sample model that is more suitable for streaming online tasks to strengthen the active learning model,and verify its feasibility in aircraft model recognition through experiments.The proposed model meets the needs of intelligent air traffic management and has broad application prospects.