Ensemble Method for Reinforcement Learning Algorithms Based on Hierarchy

Abstract

The article proposes an ensemble method for reinforcement learning algorithms. The proposed approach is on average more efficient than each of the algorithms in the ensemble separately. The article discusses the implementation of the method, which includes an ensemble of REDQ and SAC algorithms. The output from the ensemble is the output of the algorithm selected following the output of the DQN acting as the control algorithm. It is possible to ensemble other algorithms in a different quantity. Reinforcement learning is a promising area in machine learning. An important unsolved problem of reinforcement learning is the generalization of complex problems, and their solution using meta-algorithms. The proposed method can be used in complex problems consisting of many subtasks, effective solutions for which can be offered by various algorithms from the ensemble.