报告名称：Neural Control with Certificates for Safe Autonomy

报告时间：5月13日10:00

报告地点：中北校区理科大楼A228

报告摘要:

Learning-based methods, such as reinforcement learning, are receiving increasing attention for solving challenging control tasks. However, the lack of safety assurances about learned controllers poses a significant barrier to their practical deployment. In this talk, we will present a framework for learning and/or formally verifying neural controllers. Given a specification of interest, the framework jointly learns and formally verifies a controller together with a formal certificate of the specification being satisfied, both parametrized as neural networks.  The framework is applicable to stochastic dynamical systems, thus also taking into account environment uncertainty. Certificates are supermartingale-like objects that can be effectively used to formally reason about stochastic systems in a fully automated fashion. We will show how the framework can be applied to solve a wide range of challenging control tasks.

报告人简介:

Djordje Zikelic is an assistant professor at the School of Computing and Information Systems, Singapore Management University. Prior to joining SMU, he obtained his PhD in computer science at the Institute of Science and Technology Austria (ISTA) and bachelor’s and master’s degrees in mathematics at the University of Cambridge. His research focuses on formal verification and synthesis of software and AI systems and lies at the interplay of formal methods, trustworthy AI and programming languages research. His work is regularly published at premier venues in formal methods, programming languages, AI and ML, including CAV, FM, PLDI, POPL, OOPSLA, AAAI, NeurIPS and IJCAI. For his PhD work, he received the Outstanding PhD Thesis award and the Outstanding Scientific Achievement award at ISTA.