변분 추론의 증거 하한을 사용한 강화학습

Abstract

This paper presents a practical method of designing reinforcement learning (RL) algorithms with the evidence lower bound (ELBO) of variational inference (VI). The proposed approach provides opportunities to easily employ the existing results of supervised and unsupervised learning for reinforcement learning. By linking the likelihood functions with the state-action-value functions reasonably, we design the machine learning algorithms in a unified frame. As a special application of ELBO-based RL algorithms, network sparsification is introduced, which is achieved by employing a sparsity-induced regularization term. To help the overall understanding and gain physical insights, a schematic view is provided. A quadrotor is then employed to validate the proposed method.