End-to-End Optimization and Learning for Multiagent Ensembles

Multiagent ensemble learning is an important class of algorithms aimed at creating accurate and robust machine learning models by combining predictions from individual agents. A key challenge for the design of these models is to create effective rules to combine individual predictions for any particular input sample. This paper addresses this challenge and proposes a unique integration of constrained optimization and learning to derive specialized consensus rules to compose accurate predictions from a pretrained ensemble. The resulting strategy, called end-to-end Multiagent ensemble Learning (e2e-MEL), learns to select appropriate predictors to combine for a particular input sample. The paper shows how to derive the ensemble learning task into a differentiable selection program which is trained end-to-end within the ensemble learning model. Results over standard benchmarks demonstrate the ability of e2e-MEL to substantially outperform conventional consensus rules in a variety of settings.