The great Bob Fourer tweeted about an interesting article in Analytics by Prof. Vijay Mehrotra:
Columnist for Analytics seems down on PhD study in OR generally and optimization specifically http://t.co/GFZft9K4o8
— Bob Fourer (@4er) May 9, 2014
Prof. Mehrotra is somewhat down on operations research [OR] graduate programs, or more accurately, down on students attending OR programs with the intention of entering the business world. Though I agree with the sentiment and many of his points, I disagree with Prof. Mehrotra’s well-stated and well-intentioned conclusions. I’d like to explain why. His concerns are more-or-less as follows:
- Business analysts need to be able to acquire and manipulate data, and OR programs don’t teach how to do that.
- Good answers quickly are better than those provided by OR.
- You can’t do optimization without first doing forecasting.
- Business analysts work on a team, and working as a team is not taught in OR programs.
Prof. Mehrotra concludes that better choices for such students are to get an MS in analytics, or a PhD in another domain such as biology, physics, or psychology. This is like saying that it’s a bad idea to get a BS in Computer Science if you want to be web developer. You don’t need a BS to be an outstanding web developer, but don’t tell me that it’s a bad idea. If you get a BS in Computer Science and don’t pick up some web development, your CS program stinks. Similarly, if you don’t learn something about data acquisition, time-quality tradeoffs, and other kinds of analytics, your OR program stinks. And if a school’s OR program stinks, its analytics program probably stinks too. The whole thing stinks.
Prof. Mehrotra seems to have in mind one particular type of operations research graduate experience: the “lone wolf” working with an advisor on convergence proofs of sophisticated-yet-esoteric operations research algorithms, crappily implemented in Matlab. (If near a cultured urban area, Python. If near a major employer or government agency, a decomposition algorithm run on a proprietary data set.) If this kind of operations research education what he has in mind, I lament with him. Too many operations research graduate students waste their time on problems that are unworthy of their talents, and I have seen several flounder in the business world after graduating. Like Prof. Mehrotra, I would never stand in the way of someone who wants to do fundamental research in operations research methods, but operations research is so much more. We know this because the very roots of operations research lie in wartime Europe, where the very first practitioners worked in teams with data sets that were the “big data” of the time, based on forecasted data. The plans of Blackett’s Circus were not always perfect (read more here), but they were timely. They had to be. I realize it’s no longer the 1940s, but it’s important to remember that while the modern field of analytics is bigger than operations research, operations researchers were the first data scientists, and today there are countless operations researchers in industry doing great work.
Regrettably the academic view of operations research has far too often strayed from these heroic and practical roots, focusing on algorithms at the expense of modeling, individual analysis at the expense of collaborative effort, and exact solutions at the expense of useful ones. Mehrotra is right to call the operations research community out on this, but these flaws are not endemic. I was fortunate to have professors at the University of Iowa with a balanced view; Ken Atkinson had the famous Hamming quote “the purpose of computing is insight, not numbers” taped lovingly to his door. Many other past and present OR graduate students share similar experiences. My advice to a prospective student would be to find operations research programs that embrace the fullness of the field, and reject those who focus exclusively on the third decimal place in an approximation algorithm bound, or in the exponent of the worst case computational complexity for an algorithm. Moreover, when you think about it, aren’t the problems Mehrotra cites common ones for the graduate school experience in general? Doctors and lawyers will tell you that their professional school experience was a necessary but insufficient step in their career preparation. Internships, group projects, and case studies are great, but nobody pretends that these completely prepare one for the realities of professional life. You have to get your hands dirty. So it is with analytics. In the real world you find out very quickly that heuristics are not a dirty word and it’s okay to jack up the solver tolerance a few decimal places. If you know what you’re doing. This is really not any different from any other profession.
Graduate training in operations research can actually be wonderful preparation for the meaningful (and profitable) application of quantitative skills in the business world. Let’s remember a key fact: operations research is analytics. In fact, it’s the coolest, highest and purest form of analytics because it tells you what you should do. Business is about making decisions. Operations research gives them to you.
Those trained in the art can be like Marines or the mother-blanking A-Team or something. They will go where others will not. I previously led an analytics development team at Nielsen, with several recent PhD graduates with operations research training. Initially I had them focus on optimization-related projects, for example optimized media planning based on the results of predictive marketing mix models. In time I realized that my operations research experts were often the best prepared to take on new analytics projects, even predictive and (gasp!) descriptive ones. This was because they possessed a winning combination of school-taught skills and on-the-job skills. In their graduate programs they developed a disciplined approach to modeling and rock solid knowledge of applied mathematics, statistics and optimization. On the job they developed an understanding of the business domain, data acquisition, and how to make good tradeoffs between quality, time, and complexity. Not everyone who can memorize the KKT conditions can do a Nielsen data pull, it’s true. But they’ve got a better shot than most…and then they can do something cool with the data once they have it.
In summary, if students like Mehrotra’s asked me my advice, I’d ask them to think about what type of job they’d like to have. You don’t need operations research training to be successful in analytics, but it can be a tremendous asset. I would never discourage someone who takes an interest in optimization to dump it for fear of crippling a data science career. Operations research has a proud and fascinating history because it is awesome and because it is useful. Let us think about how we improve how operations research is taught so that it remains thus.