In recent years, there has been a lot of talk about artificial intelligence (AI), primarily machine learning (AA, or machine learning, ML). And yet AA has its roots (among others) in a discipline that is relatively unknown to the general public, but often to AI experts as well: Operational Research (OR).
New interest in artificial intelligence
Since 2010, artificial intelligence (AI) has become a new threat. Not just any AI technology, especially machine learning (ML).
The reason for this renewed interest is mainly due to the advent of Deep Learning (DL, Deep Neural Network or Deep Learning), a strategy that was already developed in the 1960s. It has had its ups and downs, but since the 1990’s, the results have been significant. In particular, the development and use of backpropagation algorithms (see detailed history of this algorithm here and here) has made it possible to train larger neural networks. And then of course, the advent of the GPU and so on.
Fast forward to 2010 and 2020 and now it looks like ML and DL in particular will solve everything (see Hinton’s excellent claim here). And indeed, corporations and states are literally pouring billions of dollars into the development of AI, or more specifically, ML.
One AI train can hide another …
When it comes to data science, big data, and artificial intelligence, most people have this image in mind: at least one important player is missing: Operation Research (OR). A more realistic picture would be:
In fact, all of these areas are somehow based on OR. But what is operation research?
Operations research is a field of applied mathematics that can be described as the science of optimization (and in fact much more than that). If you’re interested (and if you consider yourself a serious data scientist, you really should be!), Check out the Wikipedia page.
In fact, if you are thinking of optimizing, you should really think about OR. We will come back to this later in another post. But it is not surprising that mathematical optimization plays an important role in all analytical methods.
Now a third AI train is coming! It doesn’t ignore AA or OR, it unites them! It combines the strengths of these two fields and can compensate for the weakness of one with the strength of the other. In Funartech, the world’s pioneer in this combination is called “Hybridization of ML and OR“More and more people, companies and organizations are exploring this path. In Quebec, Ivado calls it Digital Intelligence (IN). We’ve come a long way. I remember 5 or 6 years ago when I started advocating for a combination of the two, I rarely saw People still understand what it means. Even today there are many people who do not understand or realize that uniting is not only meaningful, but in fact our most powerful method today (today, but not necessarily tomorrow!) One idea to remember is hybridization. (Used) several domains together, which, regardless of the problem, allows us to go further to solve complex (industrial) problems. However, there are more credible cases to demonstrate, not to mention that this combination (actually at least 4 The type ML / OR combination) makes it possible to do something that is impossible only in ML or OR.
Operational research, a newcomer, really?
If OR is so important, how come so few people hear it? How is this possible?
A few years ago, ML was considered a subfield of OR. In fact, ML uses OR to optimize its predictions. 5-10 years ago, it wasn’t even a debate and (practically) everyone accepted this paternity. For this reason, for example, Professor Joshua Bengio is a professor at the Department of Computer Science and Operational Research, DIRO, University of Montreal.
Personally, I always see things this way. I don’t want to reduce ML to use OR to optimize its predictions, because it’s much more than ML, but the basic idea of developing algorithms is to determine what to do for yourself (that is not to follow if-then-else rules) Do (even better, everyone codes the algorithm with the same computer language …) comes from OR and is not specific to ML or even OR.
Or a well-established field that has seen great success since World War II. It is still unmatched in terms of optimization (not even by quantum computing, see for example overcoming the challenge of quantum optimization). In another post, I will explain why ML in general is really bad for optimization. ML can be seen as a breakthrough from OR and is certainly an indication that optimization is a powerful engine for finding solutions to complex problems.