The 3 Percent Solution
Judith Norback is not a theorist, but is an expert on communication skills—especially oral presentations. She is a faculty member in the ISyE department at Georgia Tech, where she runs the communications laboratory.
Today I want to talk about about the role that theory can and should play in practical problems.
I was recently invited as her guest to attend ISyE’s “best presentations,” which raised a question about the role of theory in my mind. I should also add, for full-disclosure—that Judith is my wife and I am her husband.
The 3% Solution
The name ISyE stands for Industrial and Systems Engineering. This department is currently the best, of its kind, in the nation and has been best for over a decade, which is impressive since the competition includes departments from most of the top engineering schools.
ISyE, as most departments do, has a “capstone” course. In this course teams of seniors work with companies, applying their industrial engineering knowledge to solve real problems. This year, one team, for example, worked on improving the routing of auto parts for Honda; another on prediction of the future demand for wind turbines for GE.
The other night I went to the final presentations of the three best teams: its an American Idol type show. Each team makes a presentation of their work, there is a panel of judges, and a winner is picked. The winning team does not get a recording contract, but the sense of competition was still high. The teams take their presentations seriously, very seriously, and it showed. The presentations were clear, interesting, and professional. I wish my presentations were half as good.
What struck me about the presentations is they made an implicit point that we in theory usually avoid: Constants matter—sometimes a great deal.
Each of the teams presented work that went like this:
Our company has a real problem X that is We collected data on the problem, studied what the company does now to solve X, and determined the cost of their current solution. Then, we applied our industrial engineering methods and found a new solution S to their problem. We realized that we could use randomization, local 2-opt search, and tabu methods to make our solution S more efficient. The result was a cost savings of dollars over the previous methods.
The surprise to me was that their solutions S were often slight improvements to the current practice. In one case S was better by 3%, and in another case, by 1%. These are relatively small improvements. The cool point is the 3% was of about 190 million dollars. Thus, the savings to their company was potentially more than 6 million dollars. This is a huge amount of money, since it goes right to the bottom line of the company’s profit. I do not have an MBA, but to get 6 million dollars to the bottom line, would require 100s of millions of dollars in new revenue.
As theorists, in our thinking, in our classes, in our papers, and at our conferences, we all work in O-notation. We prove approximations, we decrease a bound from one polynomial to a smaller one, and so on. But, what I realized the other night is that there is something missing from this model. We need to have papers that show even small percent increases—provided the percent is of a large enough amount.
Thus, the question that I have been thinking about is: how can we do what the ISyE teams did? Can we apply our algorithmic tricks to get small percent increases that really matter? Would papers on this type of work be accepted at our top conferences? Would NSF fund this type of research? Are there models of how we can help in this way?
Should we work on such results? How can we help apply some of our theory “tricks” to make small percent improvements, that really matter?