Traditional active learning strategies usually focus on either uncertainty or diversity by querying the most difficult or representative sample, but fail to take into consideration the interaction between labeled and unlabeled samples. In this work, we propose a novel active learning algorithm which aims to maximize the overall performance gain of the model by considering the interaction between labeled and unlabeled samples. Our method estimates this by simulating one-step gradient descent and parameter updating on current model, and iteratively selects batch samples that maximize the expected model loss reduction(EMLR) upon data distribution. We further compared our method with a series of competitive methods on various tasks from different fields. The experimental results show that EMLR performs consistently better or equal compared to all competitors in all tasks. Extra analysis experiments have demonstrated that our method can archive a good trade-off between uncertainty and diversity.