Dynamic Ensemble Selection with Reinforcement Learning

Lihua Liu, Jibing Wu, Xuan Li, Hongbin Huang

Abstract: In this work, we propose a novel approach to ensemble learning, referred to as Dynamic Ensemble Selection using Reinforcement Learning. Traditional ensemble learning methods rely on static combinations of base models, which may not be optimal for diverse inputs and contexts. Our proposed method addresses this limitation by dynamically selecting the most appropriate ensemble member based on the current input and context, utilizing reinforcement learning algorithms. We formulate the ensemble member selection problem as a Markov Decision Process and employ Q-learning to learn a selection policy. The learned policy is then used to adaptively choose the best ensemble member for a given input, potentially improving the overall performance of the ensemble learning system. The proposed method demonstrates the potential for increased accuracy and robustness in various learning data sets.

External IDs:dblp:conf/icic/LiuWLH23