Combining Operations Research (OR) and Machine Learning (ML)

If you are an active resident of analytics land, you know that Artificial Intelligence (AI) and Machine Learning (ML) tools are the new bosses in town. Every tool, technology, and technology solution around you tries to incorporate them in their solution in some form. And all this limelight on AI and ML has pushed the classic analytics professionals like statisticians (who stood their ground and decided to stay with that title rather than getting “rebranded” as data scientists) and good old Operations Research (OR) professionals into a separate categories. Whether talking to executives, reading books, or doing secondary research, a consistent theme is that we have classified the community of advanced analytics professionals in the supply chain world into two primary categories: OR and ML professionals. But do they need to always be in two distinct buckets in the supply chain? This article discusses how OR and ML algorithms can be leveraged in tandem to address critical challenges in the supply chain world.

A brief history of Operations Research (OR)

For starters, Operations Research (OR) is an applied mathematics discipline that leverages multiple algorithms and techniques like simulation, modeling, queuing, and other stochastic and probabilistic methods to optimize or improve a business process. The foundation for this discipline was laid during World War II, and the key driver was to leverage applied math to find solutions to strategic operations and logistics problems. The discipline has evolved since then, and OR techniques are used widely in non-military scenarios today. Operations Research has been employed in logistics Engineering for quite some time now, some of the classic applications being optimizing flows, identifying optimal locations, minimizing transportation costs, optimizing transportation assets, inventory optimization, etc.

And then, new analytical approaches started evolving.

While Operations Research evolved over the years to find applications outside the military and became widely used in Industries, other fields also evolved concurrently. An example is Management Science, which essentially leveraged applied math to solve business and economics problems. And then, with advancements in computing technology, data mining tools, and database technologies,  we saw the emergence of new fields, ML, deep learning, NLP, and reinforcement learning, among many others. ML  tools quickly grew in popularity as predictive analytics tools due to their wide gamut of applications made possible by today’s technology. And then, as they grew popular, we started thinking of them as tools that generally work separately from OR tools and are helpful in different contexts and scenarios.

What was the thinking behind drawing the line between OR and ML algorithms?

In my opinion, our classic (and archaic) way of thinking has not only created the divide between OR and ML. Still, it has also thwarted opportunities to break what I call “analytics silos”.  OR and ML tools are leveraged in silos, but the critical question that we need to ask is – do they need to exist in silos?

If we think about this from a supply chain perspective, the conventional OR approaches’ challenge was that they were essentially “one-off” modeling exercises in prescriptive analytics. The challenge or drawback of this approach is that today’s supply chain and manufacturing world has become much more dynamic and complex. Many variables might change when you take recommendations from a prescriptive model like this one to the floor.

But one of the key bottlenecks in using these optimization algorithms “live” was that in the real world, where input variables, decision factors, and constraints are insanely complex, the tool would not be able to spit out recommendations as fast as shop floor operations planning would like it to. Imagine a production schedule optimization model, for example, that decides what parts should be processed together to minimize aspects like setup time, total production time, etc. Running this model “live” to plan daily production runs may not be feasible since the run time (even with today’s computing power) will be too long to be helpful as a live planning tool for complex SKU portfolios and manufacturing operations.

And this is what we generally cite as a drawback for OR. Machine Learning (ML) algorithms, when used in the true sense (i.e, used for scenarios where they should be used), can be “live” learning algorithms, but they are more geared towards prediction scenarios. And hence, we see OR and ML as two distinct buckets.

But the real magic happens when you combine these two!

Did we think about it the wrong way?

Let us revisit our production planning example. I indicated that one of the challenges these models run into is the long run time, making them unsuitable for live planning tools.

But here is another scenario. An optimization algorithm runs for days, and we build a massive data set of all input variables and output recommendations. We run these scenarios millions of times to create a large data set. THEN…. we feed this data to an ML algorithm(which can also be a Deep Learning algorithm). So now, we have trained an algorithm that can predict, based on input variables, what the expected output will be. And then, we can build a drag-and-drop interface around this ML algorithm that operators can use to select the optimal mix of parts that optimizes the production process aspects. And voila…we have created a solution combining optimization and ML algorithms.

This is just one example, and there can be many opportunities. Supply chain resiliency and agility have emerged as critical imperatives in all three of our 2022 supply chain reports, Supply Chain Planning in The Cloud, Process Automation in Supply Chain, and Inventory Management and Optimization. Resiliency and agility have emerged as top imperatives for our upcoming research report focused on data and analytics in supply chain as well. And this is another area where combining operations research methodologies with ML algorithms can make a significant impact.

We use resiliency and agility in tandem most of the time. However, the fact is that building too much agility can hamper resiliency. A good example is the Just-In-Time (JIT) inventory approach. As the pandemic highlighted, supply chains built on the premises of fast, efficient and timely synchronization, within minimal inventory being held across the network, suffered.

The planning approach of Just-In-Case, a resiliency approach leveraged by many companies these days, is not optimal either. There is a need to balance and combining OR with ML can help in this areas as well. Figure 1 shows the approach that can be taken to redefine a process designed exclusively for JIT to insert resiliency in it.

Figure 1: Evolving from JIT to resilient JIT

If you think about the three Rs of resiliency highlighted in illustration one, each of these can leverage a combination of OR and ML algorithms, to help you processes that balance agility and resiliency.

Recognize: Combining ML based pattern recognition approaches with stochastic optimization approaches, leveraging both historical data as well as real-time data being streamed, can help you identify a potential disruption sooner (and in some cases, predict them).

Respond and Reframe: This is where combining these two approaches becomes much more powerful. You can actually leverage supply chain network design solutions, that use conventional OR algorithms, on data that has been generated by ML algorithms to design a network that takes into account every possible (predictable) disruption, at different magnitudes of impact, and generate scenario recommendations for them.

The above two are a sample of the opportunities that supply chain and operations presents to combine these two powerful approaches, to help develop supply chains aligned to the current challenges.

What does this mean for SAPinsiders ?

• Enterprise analytics is not just about running multiple analytical tools in silos. It is about finding opportunities to play around. Use your analytics canvas to draw beautiful solutions by combining different analytical tools to create robust solutions.
• The suggested scenario is just one example, but you can combine these approaches across the end-to-end supply chain.
• Create analytics teams not divided by expertise areas (like OR and ML). If they are, run workshops and sessions to help them think about opportunities at a high level. Push them to evolve as analytics artists vs. modelers, programmers, and data scientists. Designations do not do any magic; skills do
• SAP and SAP ecosystem partners provide many tools you can leverage to create such innovative solutions. You can choose to build solutions from scratch using open-source tools.
• Third-party Enterprise AI tools that provide these “combined” capabilities in off-the-shelf products are also available in the market.