From October 16–19, the 2022 INFORMS (Institute for Operations Research and the Management Sciences) annual meeting was held in Indianapolis. SigOpt last attended this meeting in-person in 2019, 2021, and attended this year as well.
Summary
At SigOpt, we have observed that users sometimes care about having distinct outcomes. To address this challenge, we have been investigating the importance of diversity when performing experimentation, whether it be materials science or hyperparameter optimization. The culmination of this investigation was constraint active search, which prioritizes diversity over optimality; this was presented at last year’s INFORMS.
This year at INFORMS we presented an extension of our work on constraint active search, which uses a different measure of diversity than before (in metric space, rather than parameter space). This work will also be presented and published at this year’s Winter Simulation Conference (WinterSim) 2022, and we hope to see some of you there!
Constraint Active Search: Achieving Diversity in Multiobjective Experimentation
Authors: Eric H Lee, Bolong Cheng, Michael McCourt
Speaker: Eric H Lee
In this talk, we summarize constraint active search (CAS), which attempts to sample a diverse set of points from an unknown feasible region implicitly defined by a set of black-box constraints. CAS applies to many problems in engineering design and simulation which require balancing competing objectives under the presence of uncertainty. Sample-efficient multiobjective optimization methods focus on the objective function values in metric space and ignore the sampling behavior of the design configurations, in which sample diversity may be an important concern during experimentation. To emphasize this point, we highlight the importance of achieving sample diversity through CAS using a real-world application in additive manufacturing. We then showcase an extension of CAS to metric space using a new acquisition function we developed called the likelihood of metric satisfaction.
Bayesian Optimization Dinner
In line with tradition, SigOpt hosted one of its annual Bayesian optimization dinners this year at INFORMS and invited a number of researchers in the field to attend. Attendees ranged from researchers working at Meta and Google Brain, to faculty members at Cornell, Arizona State University, NYU, and Georgia Tech, to various students and postdocs.
INFORMS Takeaways
In prior years, Bayesian optimization was not a well-represented topic at INFORMS, which traditionally serves the operations research and applied mathematics communities, rather than the machine learning community. However, the presence of Bayesian optimization is growing rapidly at INFORMS. There were no less than four sessions on Bayesian optimization alone (contrasting with only one session in 2019 and 2021), and many, many more concerning global optimization under uncertainty. We expect the Bayesian optimization community’s representation to keep increasing in the following years.
More generally, the presence of machine learning at INFORMS is growing very rapidly, which indicates that much of the prior hesitation in the operations research and applied mathematics communities at adopting machine learning is disappearing. We will see if this prediction holds true when we see you next year at INFORMS 2023!