Combining multiple correlated reward and shaping signals by measuring confidence

Abstract

This paper falls in a line of research that investigates how reinforcement learning algorithms can be sped up by combining two AI techniques: multi-objectivization and ensemble techniques. Multi-objectivization is a process that turns a single-objective problem into a multi-objective one, essentially creating a diverse set of feedback signals for the original single-objective task. These signals can be strategically combined using ensemble techniques to speed up and improve learning. In this paper, we present a novel ensemble technique that detects where and when each of the signals is most informative, and proceeds to use that one for action selection.