Next Wednesday, 9th of April, at 14:00 (Paris time), CEDAR and Prof. Yanlei Diao will host Prof. Peter Haas from the University of Massachusetts Amherst. We invite you to join us for the talk of Prof. Haas, titled: “Graph and Generative Neural Networks for Accelerating Simulation-based Decision Making Under Uncertainty”
The event is co-organized by BNP Paribas and the École Polytechnique.
Abstract:
Stochastic discrete event simulation has long been used to support decision making for complex systems in the face of uncertainty, with applications in banking & finance, healthcare, manufacturing, defense, logistics, and many more. A key challenge, however, is that large, complex stochastic simulations can take a very long time to execute. Simulation metamodeling is crucial for enabling what-if analysis and simulation-based optimization in complex, uncertain operational settings, where results are needed quickly. A simulation metamodel is trained offline to statistically map simulation inputs to outputs; the deployed metamodel then returns online results orders of magnitude faster compared to actually running the simulation. We enhance prior simulation metamodeling in two important ways. First, we use graph neural networks to allow the graphical structure of a simulation model to be treated as a numerical metamodel input that can be varied along with real-valued and integer-ordered inputs. Second, we combine GrNNs with generative neural network components so that a metamodel can rapidly produce not only a summary statistic like E[Y], but also a sequence of i.i.d. samples of Y or even a stochastic process that mimics dynamic simulation outputs. Thus we can approximate a range of performance measures using a single metamodel and potentially use trained metamodels in digital twin settings. We then provide a novel method, called HiLo, for efficiently training a metamodel; HiLo is customized for the simulation setting and outperforms generic active-learning methods. Finally, we provide techniques for speeding up simulation-based optimization by using the metamodel, rather than expensive simulation, to compute objective values. Of particular interest are hybrid discrete-continuous optimization problems that naturally arise on our setting but have received relatively little attention. We provide a heuristic solution method based on an extension of Monte Carlo tree search and an exact method based on mixed-integer linear programming with a customized solver.
Biography:
Peter Haas joined the UMass faculty as a Professor of Information and Computer Sciences and Adjunct Professor of Mechanical and Industrial Engineering in 2017 after 30 years at IBM Research — where he rose to Principal Research Staff Member — and 20 years as a Consulting Professor in Management Science and Engineering at Stanford University. His research lies at the interface of information management, applied probability, statistics, machine learning, and computer simulation. He is a Fellow of both ACM and INFORMS and has received awards from IBM and both the Simulation and Computer Science communities, including an IBM Research Outstanding Innovation Award, an ACM SIGMOD 10-year Best Paper Award for his work on sampling-based exploration of massive datasets, and the INFORMS Simulation Society Outstanding Publication Award for his 2002 Springer monograph on Stochastic Petri Nets. Current research topics include in-database decision support, machine learning (ML) for simulation, and predictive maintenance of ML models under data drift. He is the author of over 160 conference publications, journal articles, and books, and has been granted over 30 patents, leading to his designation as an IBM Master Inventor. His work has been incorporated into IBM products and, more recently, the Apache DataSketches library.
The talk will take place in amphithéâtre Sophie Germain at the Alan Turing building.
We look forward to seeing everyone there!
