ENBIS-QSR webinar: Enhancing Robustness in Reinforcement Learning: Domain Knowledge Integration and Lower-Bound Certification

Thursday 2 Apr 2026, 16:00 → 17:00 Europe/Amsterdam

Bo Shen (New Jersey Institute of Technology)

ENBIS-QSR webinar: Enhancing Robustness in Reinforcement Learning: Domain Knowledge Integration and Lower-Bound Certification

Speaker: Yisha Xiang (Industrial & Systems Engineering Department at the University of Houston)

Chair: Bo Shen (New Jersey Institute of Technology)

Date: 2nd April 2026, at 16:00-17:00 CEST/ 10:00-11:00 EST

Reinforcement learning (RL) has become a powerful framework for solving sequential decision-making problems, particularly in complex and dynamic environments where traditional optimization techniques often struggle. This work develops novel RL-based methodologies to address critical challenges in domain knowledge incorporation and robustness certification in RL, contributing to both theoretical advancements and practical applications. We first develop a domain knowledge-informed deep reinforcement learning method to leverage prior knowledge and overcome slow convergence in conventional Q-learning. Theoretical results guarantee convergence for small-scale problems, while computational experiments on larger problems demonstrate superior efficiency and reward performance compared to traditional methods. The second part of this work proposes a novel certification framework for evaluating the robustness of RL policies under bounded adversarial state perturbations. We formulate the lower bound certification problem under p-norm-bounded perturbations as a convex optimization problem through a phi-divergence-based relaxation, and derive its dual formulation to establish tractable, risk-aware lower bounds on the expected exponential utility of smoothed policies. We further develop an empirical method to improve the certified lower bounds of RL policies. Our results show that risk-averse training generally results in policies with higher certified lower bounds than risk-neutral training, especially under larger perturbation budgets.

 

Bio:

Dr. Yisha Xiang is an Associate Professor and the Scott T. Poage Fellow in the Industrial & Systems Engineering Department at the University of Houston. Her current research and teaching interests involve data-driven decision-making under uncertainty and statistical machine learning. Her research has been funded by the National Science Foundation, including a CAREER grant, and industry. She has published articles in refereed journals, such as INFORMS journal on Computing, IISE Transactions, European Journal of Operational Research, and Naval Research Logistics. She was the recipient of the P.K. McElroy award, Stan Oftshun award, and Doug Ogden award for best papers at the Reliability and Maintainability Symposium. Dr. Xiang received her B.S. in Industrial Engineering from Nanjing University of Aero. & Astro., China, and M.S. and Ph.D. in Industrial Engineering from University of Arkansas. She serves as an Associate Editor for IISE Transactions, INFORMS Journal on Data Science, and IEEE Transactions on Reliability, and was previously on the editorial team of IEEE Transactions on Automation Science and Engineering. She has served as President of the IISE Quality Control and Reliability Engineering Division and Chair of the INFORMS Quality, Statistics, and Reliability Section.  She is a senior member of INFORMS and a member of IISE.

 

 

16:00 → 17:00 ENBIS-QSR webinar: Enhancing Robustness in Reinforcement Learning: Domain Knowledge Integration and Lower-Bound Certification

Speaker: Yisha Xiang (Industrial & Systems Engineering Department at the University of Houston)