Machine Learning: Is it really the hero that the Operations Research community needs?

Siddhartha Nambiar

by Siddhartha Nambiar
Ph.D. candidate in Industrial Engineering, North Carolina State University

About a year ago, I was presenting a poster outlining my current research at a conference. My work was about an application of Markov Decision Processes, and I had had several people stop by at my poster to interact with me. Most of the questions that came my way were easy to deal with, until one individual asked me something that motivated me to write this article. His question was simple – why not just use machine learning to do this?

I’m not going to tell you how I responded to him because that would require I outline parts of my research which is outside the scope of this article. What I will tell you is that it got me thinking about how Machine Learning as a buzzword is used in so many different contexts today. A fellow student once mentioned to me that he felt graduate students in Operations Research (OR) and Statistics are at a substantial handicap when compared to graduate students in Machine Learning (ML), despite being in substantially overlapping subjects. While I wasn’t sure if I agreed with the first part of his statement, the second part resonated with me. Both ML and OR share foundations in probability theory, optimization, and linear algebra.  There is little doubt that both OR and ML possess their own independent domains. OR, as the application of scientific and mathematical methods to the analysis of complex systems, was invented during World War II. ML, as the branch of computer science that gives computers the ability to learn without being explicitly programmed, is more recent and has been around since the 80s.

Since its inception, however, ML has been slow to catch on. Comparing the interest between the two fields shows ML only break away after 2010. Perhaps advances in computational capabilities played an important role in the rise in popularity of Machine Learning. The interesting question then becomes – why has this not extended to Operations Research? The perceptions among practitioners about what the two topics are and how they contribute to the research community are variable. As one researcher at Google explains, perhaps the difference between the two fields is just semantics. His point is that, at the end of the day, people are people and could as easily work under the umbrella of OR as ML. However, he admits that some ML algorithms sometimes tend to be ‘hacks’; they rely on a ton of methods that depend on intuition rather than theory. 

A paper published in the Journal of Machine Learning Research in 2006 (Bennet and Parrado-Hernandez, 2006) looked  at how the two fields are largely intertwined. The authors state that ML researchers have embraced the advances in optimization thus allowing new types of models to be pursued.

The natural thought that follows then, is that the difference between the two fields is not as rooted in theory as it is in semantics and accessibility, the latter being an important factor. Today, a very large number of people recognize the term Machine Learning. While this is largely in part due to our world being overrun by tech companies that use the term as a means of catching the public’s attention, nuances in the way that ML is being taught cannot be ignored. As it stands today, ML is not something that too many people understand. When I think about the individual who suggested I apply ML to my research at the conference, I can’t blame him. The world we live in is more about ‘using ML’ rather than ‘learning ML’. In other words, give the man a fish rather than teach him to fish, because teaching him to fish would involve teaching him multi-dimensional calculus, advanced linear algebra, optimization theory, and advanced probability theory, just to start with. Whether this is a problem or not is up to the community to decide. Proponents of the status quo will argue that this is how academic and technical research has been taking place for many years and that ML is no different. But is ML really ‘no different’, especially now that policy makers in congress are starting to pay attention to the way ML and AI are shaping the lives of the public?

In summary, Machine Learning is more popular than Operations Research because of two reasons. The first is that the biggest part of the analytic decision-making process has now been partially shifted to the machine. The investment in machine learning is a natural evolution in technology and humanity’s demand to create technologies that extend our own capabilities. And while the same can be said of advancements in Operations Research, the second reason for ML’s popularity is what really seals the deal – the fact that it works! And not just in the sense that it provides us with excellent analytic solutions to difficult problems, but in the sense that the popularity has caught on. A direct outcome of this meteoric rise in popularity is that more and more research is currently focused on improving existing methods. There is no doubt that OR benefits from this as well, since it has become increasingly coupled with ML. However, whether the semantics of OR can survive the current outburst in ML interest remains to be seen.


 Bennett, K. P., & Parrado-Hernández, E. (2006). The interplay of optimization and machine learning research. Journal of Machine Learning Research7(Jul), 1265-1281.