Research On Active Learning Algorithm Based On Reinforcement Learning

When we use traditional supervised learning methods to do classification tasks,we hope to have more annotated samples to learn.Just as the saying goes,“ the more you see,the more knowledgeable you will be ”.The more annotated samples we have,the better classification model we will get.However,the reality is that although we can get lots of samples,most of them are without labels and it often takes us plenty of time and resources to annotate them precisely.What’s more,some samples even require the annotation of experts in the field.Fortunately,the goal of active learning is to label as little data as possible without compromising on performance of models we get.There is a long history in active learning,and a lot of sample selection strategies have been produced until now.However,many of them are heuristic.In this paper,we put forward a sample selection strategy——RL-Active.It regards the selection process of active learning as a “game”,and introduces the reinforcement learning method.Under the situation of active learning,our method replaces traditional heuristic sample selection strategy with deep reinforcement learning,which makes the process of sample selection can be learned and more universal.In the part of experiment,we compare the method with random strategy and uncertainty-based sample selection strategy on a small artificial data set,and tried to combine them.Finally,through the observation of our experiment results,we analyze some strengths and limitations of the RL-Active strategy.