Intelligent trainer for Dyna-style model-based deep reinforcement learning

• Main Authors: Dong, Linsen, Li, Yuanlong, Zhou, Xin, Wen, Yonggang, Guan, Kyle
• Other Authors: School of Computer Science and Engineering
• Format: Article
• Language: English
• Published: 2022
• Subjects: Engineering::Computer science and engineering; Reinforcement Learning; Ensemble Algorithm
• Online Access: https://hdl.handle.net/10356/159633
• Institution: Nanyang Technological University

Model-based reinforcement learning (MBRL) has been proposed as a promising alternative solution to tackle the high sampling cost challenge in the canonical RL, by leveraging a system dynamics model to generate synthetic data for policy training purpose. The MBRL framework, nevertheless, is inherently limited by the convoluted process of jointly optimizing control policy, learning system dynamics, and sampling data from two sources controlled by complicated hyperparameters. As such, the training process involves overwhelmingly manual tuning and is prohibitively costly. In this research, we propose a "reinforcement on reinforcement" (RoR) architecture to decompose the convoluted tasks into two decoupled layers of RL. The inner layer is the canonical MBRL training process which is formulated as a Markov decision process, called training process environment (TPE). The outer layer serves as an RL agent, called intelligent trainer, to learn an optimal hyperparameter configuration for the inner TPE. This decomposition approach provides much-needed flexibility to implement different trainer designs, referred to "train the trainer." In our research, we propose and optimize two alternative trainer designs: 1) an unihead trainer and 2) a multihead trainer. Our proposed RoR framework is evaluated for five tasks in the OpenAI gym. Compared with three other baseline methods, our proposed intelligent trainer methods have a competitive performance in autotuning capability, with up to 56% expected sampling cost saving without knowing the best parameter configurations in advance. The proposed trainer framework can be easily extended to tasks that require costly hyperparameter tuning.