11 Things I Learned at the 2022 INFORMS Business Analytics Conference

INFORMS held their great business analytics conference last week.

Here are my 11 learnings:

One, we should go bigger and bolder. Anne Robinson of Kinaxis mentioned a Richard Larson video where he urged professors in our field to be bolder and take risks. He thinks they should be framing and developing new theories to solve new world-changing problems. I think this advice applies to data scientists in large organizations, software vendors (Kinaxis fusing machine learning and optimization), and entrepreneurs.

Two, General Motors went big and bold by reinventing the vehicle design process. This project lived up to Richard Larson’s challenge. Most product design relies on intuition and judgement. GM created sophisticated machine learning and optimization algorithms to help the designer make better choices on what features to add to new vehicles.

This project added over $2B of additional profit in 2019 and 2020. I suspect the savings and impact is much larger. This project seemed like a transformative leap for GM. And, this idea could transform how product managers, including in software, work. (See GM video here.)

Three, the practical AI community of Chile showed that innovation is also about speed. We tend to think of innovation as creative new products, theories, or techniques. But, ease-of-use and speed can also be innovative. The team from Chile was able to apply many practical AI techniques to help the government make better decisions in battling Covid. They had to come up with new ideas to get answers fast. As a result, they saved a lot of lives. They were the well deserved winners of the Edelman Prize in 2022. (See video here.)

Four, you could build a business around checking data for logical flaws. Irv Lustig gave a great talk on the importance of good data practices (with a reference to Tidy Data) An interesting part of his talk was about the need to look for logical flaws in the data.

Data can seem “clean” because it fits in tables and doesn’t violate any numerical rules. But, it can still have logical flaws the lead to problems when using it for analysis and to make decisions. For example, someone may have entered minutes when they meant hours, or show more product shipped from a warehouse than into it.

I’m sure many people already check for these flaws. But, there still seems to room for a company to make these checks more systematic and widely available. Pete Cacioppi’s ideas in Blood, Sweat, and Tears can be extended to help.

Five, Toyota creatively captures driver safety constraints. Filippo Foccaci of DecisionBrain explained how Toyota builds inbound milk-runs that minimize changes for day-to-day and month-to-month for the drivers. Toyota wants their drivers to have the same routes so they get to know them, and, therefore promote safety. Even while adhering to this, the solution helped Toyota save money, reduce the planners time from a week to a minutes, and be more responsive (like during the pandemic).

Six, I need to give Constraint Programming another look. I have mostly ignored Constraint Programming (CP) in favor of Mixed Integer Programming (MIP). In Filippo’s Toyota talk, he mentioned that CP is a great tool for quickly building local search algorithms. That is, CP is much faster when your other choice is to create a custom heuristic to find good feasible solutions. I always feel that it is better to use generic solvers wherever possible.

Seven, I need to give non-linear solvers another look. Ed Rothberg of Gurobi talked about how, in version 9.5, certain types of non-linear model could be solved fast. I think most business problems can be solved with linear models and clever modeling tricks— but not all of them. I’ve been skeptical of non-linear solvers. It sounds like I need to give them another look.

Eight, Doug Gray of Walmart gave ten reasons why data science projects fail. His talk is worth its own blog post. Here are three quick tips: make sure allow time for lots of testing, perfect is the enemy of done, and expect that getting the project into production can be up to a hundred times harder than getting the prototype working.

Nine, Amazon used optimization to give workers flexibility in picking their schedules. Scheduling e-fulfillment centers is hard. These are like factories with many people and different processes. They have the added twist of highly variable demand and a workforce that changes week to week. I have seen these projects up close, like with American Eagle.

For Amazon, to make scheduling easy, they used to have a fixed set of shift schedules. But, it didn’t give the employees flexibility— which led to more people quitting and fewer people applying. And, it didn’t help Amazon deal with their demand variability. Amazon created a portal to allow employees to select the days and shift times they preferred. The scheduling system would help find schedules that worked for everyone, giving priority to people with seniority. This system would have been impossible to scale and manage without optimization.

Ten, Talithia Williams inspired me to watch NOVA Wonders: Can we build a brain? Her spoiler was- “no.” I love attempts to demystify the hype around our progress towards Artificial General Intelligence. This looks like another good source.

Eleven, the Edelman prize is a real gem for INFORMS. In addition to General Motors and Chile, I heard that the other finalists had excellent projects.

The US Census Bureau built a large scale routing and assignment model. Alibaba used machine learning and unique optimization to improve inventory management. Merck improved biomanufacturing effectiveness. Jansen (J&J) used machine learning to predict the spread of covid to speed up their vaccine trails.

All of these finalists should be celebrated and promoted to inspire and grow INFORMS.

I could easily write many more learnings and there were many talks I couldn’t see.

I hope this conference continues to grow.