December, 8 2017
Tahir Ekin, McCoy College of Business, Texas State University
The INFORMS Annual Meeting was held in Houston, TX on October 22-25th, 2017. Initially, there were concerns about the readiness of Houston to host the conference after Hurricane Harvey. The organizing committee conducted a series of evaluations and decided to help the city get back to normalcy by holding the conference and doing some fundraising. Houston has taken a big hit, but at least the downtown areas and convention center were ready by the time of the conference. The following is my own take away from attending the conference, which may be relevant to the ISBIS community.
INFORMS Annual Meetings have traditionally brought more than 5,000 operations researchers together, having more than 70 concurrent sessions. They are becoming more relevant to business and industrial statisticians, with the recent growing emphasis on analytical and statistical methods. INFORMS now provides a professional certification for analytics practitioners, called CAP, which resulted in a higher attendance of practitioners at their annual meetings. For more details, please check the 2017 annual meeting web site.
This year, the conference had more international attendees especially because of attendance of many firms from China, highlighted by the sponsorship of Didi. The plenaries and keynotes included topics on sports scheduling, revenue management, healthcare performance improvement, space risk management, genetic algorithms, poker, oil and gas analytics and bike sharing. In the following, I will present an overview of more statistics related keynotes.
One of the plenaries I attended was given by Robert Phillips, who serves as the Director of market-place optimization data science at Uber. He talked about the challenges associated with managing both sides of a passenger-driver relationship given unknown demand and supply within a dynamic temporal-spatial marketplace. Driver positioning and pricing as well as dynamic rider pricing incentives are among the main problems Uber has to face on a daily basis. They utilize a variety of pricing and revenue management optimization models. For these models, forecasting passenger behavior is paramount for short (minute), mid (hour), and long (week) term forward looking decisions. Another main challenge is to plan for unexpected large events. They utilize classification methods such as decision tree boosters and deep learning methods. In all they do, scale is the real challenge, since they deal with 20 million prices, 5 billion forecasts and 30 million dispatch evaluations every minute.
Another plenary I want to provide an overview about was on probabilistic deep learning. In recent years, there is more emphasis on the properties of machine learning methods, and their theoretical and practical guarantees at INFORMS annual meetings. Last year, Robert Tibshirani of Stanford gave a plenary talk. This year, a plenary talk was given by Richard Baraniuk of Rice University discussing probabilistic deep learning. He discussed the big picture of the cycles between modeling and data-driven methods; and suggested that they feed off each other. Although we may be in a data-driven era, he argued that models can be beneficial and complementary by explaining where the data-driven methods may fail. Naturally, this has been of interest to statisticians. For instance, we hear a lot about deep learning being revolutionary for many applications of machine learning. Mostly, such models are treated as black box methods. Their uncertainty is not explicitly studied using the tools of probability theory. The open questions include the reasons for the effectiveness of methods based on deep learning, when and why they would fail, as well as ways to improve them in a principled fashion. Dr. Baraniuk argued that combining Bayesian approaches with deep learning can be beneficial. The main goal would be to create probabilistic generative models that can explicitly capture nuisance through latent variables. Then, we can carry out optimal inference using standard statistical tools, identify formulas that correspond to computations in deep convolutional neural network. He illustrated some of his work on visual object recognition.
Overall, I think this year’s meeting has been successful to provide a platform where current industry problems and the use of operations research/statistical methods were discussed. Next year’s meeting will be in Phoenix, Arizona in November 2018. You can view the INFORMS conference calendar using the following link: https://www.informs.org/Meetings-Conferences/Conference-Calendar