Model-Based Bayesian Reinforcement Learning in Large Structured Domains.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Reinforcement Learning

Background:

• Model-based Bayesian reinforcement learning (RL) is of significant interest for addressing the exploration-exploitation tradeoff.
• Current methods face scalability limitations due to the complexity of posterior inference over model parameters.

Purpose of the Study:

• To develop a scalable Bayesian framework for model-based reinforcement learning.
• To simultaneously learn the structure and parameters of a dynamical system and plan actions.

Main Methods:

• Utilized factored representations to manage model complexity.
• Integrated online planning techniques with Bayesian inference.
• Developed a novel Bayesian framework for joint learning and planning.

Main Results:

• The proposed framework improves the scalability of Bayesian reinforcement learning.
• Demonstrated the ability to learn dynamical system structure and parameters.
• Enabled simultaneous planning of near-optimal action sequences.

Conclusions:

• The developed Bayesian framework enhances the practical applicability of model-based RL in larger domains.
• This approach offers a unified solution for learning and planning in dynamical systems.