Problems in Statistical Learning Theory: From Graphical Models to Reinforcement Learning

Statistical machine learning is an umbrella term for the algorithms that explore and leverage the randomness in the observation to estimate the target quantities. Statistical machine learning algorithms have achieved tremendous success across a wide range of applications, including computer vision, natural language processing, and robotics. Among them, the graphical model and multi-agent reinforcement learning (MARL) are two popular topics that explore the relationship among variables or agents to improve the efficacy of algorithms. Although the techniques related to the graphical models and MARL have achieved great empirical progress, the theoretical understanding of them needs more attention.

Fengzhuo’s research focuses on designing efficient algorithms for graphical models and MARL with performance guarantees. For the graphical model problems, the requirements from the newly appearing scenarios, such as robustness and active sampling, propose new challenges for the classical passive algorithms. His work focuses on the formulation of new learning settings and the design of provably efficient structure learning algorithms for these new scenarios. For MARL, the existing algorithms cannot achieve the optimal convergence rate for a large number of categories of Markov games. His work aims to improve the performance of the existing algorithms and provide theoretical understanding of the algorithms that use neural networks.