Title：AI-Assisted Discovery of Symbolic Laws and Decision Formulas
中文题目：人工智能辅助的符号规律与决策公式发现、
时间: 2026年9月1日 上午10:30-12:00
地点:管理科研楼一楼第二教室
主讲人:Yuan Zhou (Tsinghua University)
Bio:
Yuan Zhou is an Associate Professor at the Yau Mathematical Sciences Center, Tsinghua University. He received his B.Eng. in Computer Science from Tsinghua University in 2009 and his Ph.D. in Computer Science from Carnegie Mellon University in 2014. Prior to joining Tsinghua, he was an Instructor in Applied Mathematics at the Massachusetts Institute of Technology and an Assistant Professor at the University of Illinois Urbana-Champaign and Indiana University Bloomington.
His research focuses on data-driven decision-making and artificial intelligence for science, spanning operations research, machine learning, and optimization. His work has appeared in leading journals and conferences in operations research, management science, machine learning, and theoretical computer science, including Operations Research, Management Science, Mathematics of Operations Research, Production and Operations Management, SIAM Journal on Optimization, Nature Machine Intelligence, Journal of Machine Learning Research, ICML, NeurIPS, ICLR, COLT, STOC, FOCS, and SODA. He currently serves as an Associate Editor for Operations Research and Operations Research Letters.
Discovering concise and interpretable mathematical laws from observational data while respecting domain knowledge is an important problem in AI-assisted scientific discovery. This talk will primarily present Physical End-to-End Symbolic Regression (PhyE2E), a framework for discovering mathematical formulas that describe relationships among physical variables. Symbolic regression faces several fundamental challenges, including the exponential growth of the expression space, sparse optimization signals, the difficulty of incorporating physical priors, and the tension between predictive accuracy and formula simplicity. PhyE2E incorporates physical dimensions, formula complexity, candidate operators, and constants into an end-to-end model. It learns dimensional consistency while generating formulas, decomposes complex multivariate expressions into smaller subproblems through variable splitting, and employs search algorithms for local refinement. Experiments show that PhyE2E achieves substantial improvements in symbolic accuracy, dimensional accuracy, and formula complexity.
We further apply PhyE2E to several problems in space physics. The resulting formulas characterize both short- and long-term variations in solar activity, identify a physically interpretable relationship governing near-Earth plasma pressure, and describe differential solar rotation in high-latitude regions where observational data are limited. These applications illustrate how AI can move beyond fitting observational data to assist scientists in discovering mathematical expressions that can be interpreted, tested, and investigated further.
Time permitting, I will also briefly introduce FormaTheoria, our recent work on using AI to assist the formalization and machine verification of mathematical proofs. I will conclude with an ongoing project motivated by operational decision-making. High-quality policies for complex operations problems often rely on model-specific analysis and manual derivation, whereas policies learned through neural networks can be difficult to interpret and lack rigorous performance guarantees. We are exploring the use of symbolic planning to discover concise, interpretable, and directly implementable decision formulas. We further aim to apply AI-assisted methods to verify their theoretical properties, thereby providing provable guarantees for AI-discovered decision rules.