At present, most of the research on reinforcement learning is based on an engineering approach where specific questions come first, followed by problem-specific engineering or even brute-force solutions, with fundamental theories last. But Mengdi Wang has a different approach. She starts from the basics of reinforcement learning, tries to understand their fundamental theory and complexity, develops provably efficient algorithms, and eventually applies them to real-life problems.
Wang is committed to promoting the theoretical foundation and application of reinforcement learning. The outcomes of her research can be extended to financial technology, medical artificial intelligence, robotics, and other applied areas where reinforcement learning sits at the brain of future complex systems.
After graduating from Tsinghua University in 2007, Wang went to MIT as a graduate student at age 18. Only 6 years later, she received her MS and Ph.D. in Electrical Engineering and Computer Science with a minor in Mathematics from MIT, and then joined Princeton University as an Assistant Professor in the Department of Operations Research and Financial Engineering.
Her research has resulted in accelerated optimal-complexity algorithms for a number of computation challenges, including stochastic composition optimization, non-convex sparse optimization, online dimension reduction, and Markov Decision Processes (MDP). Wang was the first person to propose multi-level stochastic gradient methods for nested composition optimization over a random path. She also created the first stochastic primal-dual method for the online solution of MDP, which could provide theoretical proof of the optimal policy and algorithm of a reinforcement learning system. Her group developed the first sample-optimal reinforcement learning algorithm, which is provably efficient in using data and learns the optimal policy on-the-fly after observing a minimal number of samples.
By combining the ideas of statistics and optimal control systems, Wang’s research group aims to integrate reinforcement learning into a complex system such as smart medical diagnosis or FinTech, creating a new perspective on solving risk management, big data analysis, and medical and financial decision making problems. Her research goal is to tackle the scalability and generalization challenges of reinforcement learning, as well as the ongoing challenge that AI is over-reliant on massive data.