Instance-Based Ensemble Selection Using Deep Reinforcement Learning

Ensemble selection is a very active research topic in machine learning area. It aims to achieve a better performance by selecting a proper subset of the original ensemble, which is essentially a searching problem in large combinatorial spaces. In this paper, we propose an instance-based reinforcement learning (IBRL) model, that selects distinct subsets for different instances. Specifically, we use deep Q-network to approximate the optimal policy. Rather than considering the overall performance of each classifier, the network learns from the feedback of classifiers on individual instance, so that it generates non-static subsets for different instances. Experiments are conducted to compare our model against state-of-the-art approaches for both selection and combination. The proposed method generates promising results and it shows exceptional advantage in large scale distributed environment. Due to the environment-free characteristic of reinforcement learning, our model is adaptable to various real world tasks with minimal changes.