Skip to main content
Open Access Publications from the University of California

UC San Diego

UC San Diego Electronic Theses and Dissertations bannerUC San Diego

Understanding Model-Based Reinforcement Learning and its Application in Safe Reinforcement Learning

  • Author(s): Hu, Dingcheng
  • Advisor(s): Gao, Sicun
  • et al.

Model-based reinforcement learning algorithms have been shown to achieve successful results on various continuous control benchmarks, but the understanding of model-based methods is limited. We try to interpret how model-based method works through novel experiments on state-of-the-art algorithms with an emphasis on the model learning part. We evaluate the role of the model learning in policy optimization and propose methods to learn a more accurate model. With a better understanding of model-based reinforcement learning, we then apply model-based methods to solve safe reinforcement learning (RL) problems with near-zero violation of hard constraints throughout training. Drawing an analogy with how humans and animals learn to perform safe actions, we break down the safe RL problem into three stages. First, we train agents in a constraint-free environment to learn a performant policy for reaching high rewards, and simultaneously learn a model of the dynamics. Second, we use model-based methods to plan safe actions and train a safeguarding policy from these actions through imitation. Finally, we propose a factored framework to train an overall policy that mixes the performant policy and the safeguarding policy. This three-step curriculum ensures near-zero violation of safety constraints at all times. As an advantage of model-based method, the sample complexity required at the second and third steps of the process is significantly lower than model-free methods and can enable online safe learning. We demonstrate the effectiveness of our methods in various continuous control problems and analyze the advantages over state-of-the-art approaches.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View