Nobel Prize Winners Still Have to Sell-- and Six Other Lessons from the INFORMS Conference
Last week, INFORMS (the professional society for Operations Research or, I would say, Practical AI) held its big annual conference. This conference has much more academic flavor than the one in the Spring, the business analytics conference.
There were hundreds of talks. I attended very few, but I still picked up some good lessons:
One, even Nobel Prize winners still have to sell.
Alvin Roth, a Nobel Prize in Economics winner for his work on market design, spoke about interesting models for matching medical residents, assigning students to schools, and kidney exchanges.
I found it more interesting that he mentioned that he had to work on selling his work. Having a Nobel Prize doesn’t make you immune.
He concluded that his work in school assignment was about “building algorithms and communicating what those algorithms do.”
He mentioned that having the perfect algorithm is not enough. You should also understand and solve the problems that may prevent your algorithm from being used.
And, to make kidney exchanges work better, he visits many countries and talks to health organizations.
Two, we’ve barely scratched the surface of parallel computing for optimization.
Jonathan Eckstein of Rutgers gave a good talk on using High-Performance Computing- HPC (basically massively parallel computing) in optimization.
His message: optimization engines can take advantage of a few cores, but nothing close to using what is available or what other fields (like Deep Learning) use.
The silver lining is that optimization engines have made big improvements over the decades, and we have breakthroughs here to look forward to.
Three, we can be involved in big strategic decisions.
I’ve often wondered if we could build more board-level models to help companies allocate capital to hit long-term strategies.
Katharina Ley Best and Jeremy Eckhause of RAND presented a model they built for the Department of Defense to do just that.
The model looked at all the possible programs the military could invest in (like more night goggles or more tanks), the usefulness of those programs, and the potential strategic needs in the future (what types of conflict and terrain might be needed).
The model would suggest how to best spend the budget— what programs to invest in and how that would help meet (or not meet) different strategic goals.
The military has a lot of data, but I’m guessing it is just as bad as corporate data. The solution allowed for quick re-runs so that the decision-makers could adjust the strategies, budget, or usefulness of each program.
I suspect we could be more models like this for companies.
Four, good stories help us sell our profession. Using OR to resettle refugees is a great example.
I’m always looking for good and relatable stories to tell about optimization. I like to use them in my class and when I’m telling people about Practical AI.
Andrew Trapp and others are building models to help resettle refugees. The models allow the agencies to find the right location for families to maximize their chances of finding employment and connections.
These models make for great material in a class and good PR.
Five, you can build models that give control to people.
Alessandro Mancuso of Kone (a global leader in elevators and escalators) talked about scheduling technicians to do check-ups, maintenance, and repairs to elevators. Instead of having the optimization engine return the best sequence, it suggested a list of possible sites over the next day or two. This allows the technicians to control their schedule (helping with retention) and provides a buffer for the many disruptions (if a person is stuck in an elevator, the plan is quickly forgotten).
This was similar to a point made by Warren Hearnes of Best Buy. He said that the worker’s view of the schedule must be considered. For example, this week’s schedule may require 12 workers at a store. But, if only ten were needed last week, the workers may feel they have more freedom to take off because there are others to cover. This has to be accounted for when schedules are built.
Six, making dress shirts is hard (or sometimes a good solution gets us to a better starting point).
Delman Lee of TAL Group (the makers of 1 in 6 dress shirts in the US) spoke about the difficult optimization problems in making shirts.
There are still optimization problems where people are better than the algorithms—like laying out the patterns on fabric to minimize waste.
In scheduling, the optimization algorithms only handle one function. But that is OK. That one function is complex, and getting that right gives the rest of the factory a good starting point.
I liked his idea that optimization algorithms can provide solid value by just getting us started. They don’t have to solve the whole problem.
Seven, successful large-scale optimization projects can take time.
Stéphane Dauzère-Pérès from Mines Saint-Etienne (an engineering graduate school in France) runs a research group that places Ph.D. students in companies. The students are placed in companies, getting access to real data, real problems, and real research challenges.
He mentioned that the best work comes out of long-term engagements.
Interestingly, a similar research group in Australia (OPTIMA), doing practical research with industry, wants companies to commit for five years.
Big optimization problems are hard. In the commercial world, companies want results fast, and vendors want to promise fast results. The truth is that some large-scale optimization projects can take time.