Unleashing Synergy: The Convergence of Artificial Intelligence and Operations Research to transform the Supply Chain industry

About two decades ago, Jack Levis set out to solve a tough problem for the delivery company he worked for. His sharp mind helped him understand the intricate challenges that an average driver faces on a typical day. Deciding in what sequence to deliver packages to three different locations is indeed a rather simple problem, implying no more than six different ways to do things. Increasing that to five locations becomes a little challenging: 120 ways. But five is simply too small a number. How about ten? Here we start to see a real challenge: 3.6 million alternatives to deliver ten packages. By the time Jack estimated the complexity involved in deciding the route for 20 different locations, it was clear that the nearly 2.5 trillion alternatives to choose from were just impossible to conceive for any human being. Enough to dismay anyone, a typical driver nowadays makes more (in some cases, many more) than 100 deliveries a day.

This was the type of problem apt to be solved through Operations Research (OR). During World War II, OR emerged as “the science of winning,” a moniker coined by the Allies to emphasize that OR represented a scientific and effective approach to a military strategy that could give them an edge over their adversaries. And an edge it certainly did give. According to some scholars, the results obtained by the early OR groups before 1942 were consistently successful and, in some cases, stunningly so. One of the most remarkable achievements was the significant enhancement of Coastal Command’s aerial assaults on German submarines, which reportedly boosted efficiency by 400% to 700%. Other operational dilemmas that received OR analysis included the integration of radar into an air defense system, procedures for night operations, the effects of weather and other factors on defensive air operations, and strategic bombing.

At the end of the war, many of the American and British scientists, mathematicians, statisticians, and engineers who had worked in those military operational research units returned to civilian life in universities and industries. Petroleum and telecommunication companies were among the first to use OR concepts. By the mid-1970s, for instance, Bell Laboratories boasted a large workforce comprising several hundred individuals, many of whom had received specialized training in operations research and related disciplines. These experts utilized mathematical models and computer-based analysis to solve operational and planning challenges. Over time, technological improvements in computational capacity accelerated the expansion of OR. Currently, it is a discipline that enjoys extensive utilization in various industries and sectors, encompassing healthcare, logistics, manufacturing, and beyond.

Convergence of Artificial Intelligence and Operations Research

In the arsenal of analytics tools, both OR and Artificial Intelligence (AI) play essential roles, demanding advanced mathematical skills while delivering exceptional competitive advantages. Their primary contrast lies in approach: AI is driven by a “data-centric” methodology, while OR takes a “decision-centric” approach.

The data-centric approach gained relevancy with the surge of “big data”, it powered the initial stages of the analytics era with tools like descriptive & inferential statistics, probability, and data mining. This methodology operates on the premise of enabling data to express itself freely, revealing insights by uncovering patterns and implicit relationships. In the more recent Analytics 4.0 era, AI and cognitive technologies are geared towards problem-solving and learning, reducing human involvement in insight generation. At the forefront of this era is machine learning (ML), where machines develop models, assess their fit to data, and iterate for improved models. Neural networks, present since the 1950s and popularized in the 1990s, are statistical versions of machine learning that apply data transformations to make inferences about different data features. Modern versions like deep learning address intricate problems by performing complex data transformations to establish a self-sufficient neural network capable of intelligent decisions. While setup time is extended, deep learning yields instant high-quality results, enabling organizations to be more nuanced in approaching customers and markets and respond swiftly to dynamic data. ML models tend to surpass human-analyzed models in accuracy due to their consideration of numerous variables and parameters in different configurations. They explore “ensemble” strategies and parameter tuning through diverse algorithm types to find the best explanatory model. An advanced step in deep learning is “adversarial networks,” enabling computers to generate new, nearly authentic data such as images.

In contrast, rather than building a model based on larger and larger data sets, the “art” of decision-centric approaches lies in identifying the crucial information from the system that effectively represents key interactions and states. This begins by understanding the required decisions and the necessary insights for enhanced outcomes. This approach necessitates domain expertise, yielding specific resource allocation recommendations under varied scenarios. Optimization, a core OR technique, employs mathematical models based on equation systems to solve problems like traffic engineering, inventory management, scheduling, networking, and other pivotal management issues. Simulation, another key OR realm, replicates system behavior to incorporate greater complexity and uncertainty, predicting performance before implementation or driving changes in existing systems. Discrete Event Simulation (DES) aptly models complex interactions within dynamic stochastic systems like queuing systems in call centers, hospitals, and more, serving as the foundation for creating digital twins. Decision analysis also holds significant importance within the realm of OR, it furnishes tools for structuring decisions and evaluating alternatives, aiding the decision-making process with mathematical forms, methods, and logical organization of decision-maker preferences. Game Theory steps in when multiple decision-makers are involved. As Decision and Management Sciences borrow quantitative methods and modeling approaches from OR to enhance organizational decision-making, the three terms are often used interchangeably. Additionally, OR, focusing on system behavior, is occasionally referred to as systems analysis.

In practice, AI and OR collaboratively harness their individual strengths, establishing a symbiotic synergy. Beyond the prevalent “predict-then-optimize” paradigm seen in various modern data applications, where predictive outputs from ML models are fed into an optimization model, integrated approaches result in significant effectiveness. Notably, non-linear optimization techniques are used in ML and neural networks to update model parameters iteratively and to train models. Gradient descent algorithms like stochastic gradient descent (SGD) and Adam exemplify this convergence, as they compute loss function gradients to iteratively refine parameters, reducing overall loss. Furthermore, simulation models can generate realistic, clean, labeled data for training AI models. ML techniques are often employed to enhance optimization processes too, making them more efficient, adaptable, and capable of handling complex and high-dimensional problems. For instance, ML models can act as surrogates for computationally expensive objective functions in optimization problems. Instead of repeatedly evaluating the actual function, the ML surrogate is used to predict the function’s behavior, significantly speeding up the optimization process. ML techniques can also tackle combinatorial optimization problems by learning patterns from data and making informed decisions about how to explore or exploit solution spaces. Additionally, ML models can help handle constraints efficiently by predicting the feasibility of solutions without explicitly evaluating each one. This aids in avoiding infeasible regions and speeding up optimization.

AI-OR Integration for Complexity Management

The integration of AI and OR presents a formidable solution for managing complexity in business operations. In complex systems, the behavior of the collective takes relevancy over the behavior of the individual. The complexity profile is the amount of information that is required to describe a system as a function of the scale of description. Typically, larger scales require fewer details and therefore smaller amounts of information. Companies benefit greatly when they can leverage vast datasets and computational power associated with data-centric analysis. However, these approaches may miss important dynamics present at multiple scales which cannot necessarily be revealed by data only. By integrating the contextual knowledge and subject matter expertise contributed by decision-centric analysis companies can more easily determine which information is pivotal. AI’s data processing and predictive capabilities and OR’s prowess at applying the right decision framework to describe dynamic business systems give enterprises a comprehensive toolset to address complex challenges head-on. As an illustration, consider supply chain management, where the synergy of AI and OR-driven algorithms can yield a competitive advantage by simultaneously optimizing inventory levels, transportation routes, and demand forecasting. This integration creates a harmonious and streamlined system. Also, in the context of supply chain operations, the convergence of OR and AI to formulate digital twins is gaining traction and presents substantial opportunities. This alliance empowers enterprises to simulate diverse scenarios, predict disruptions, fine-tune operations, and enhance decision-making accuracy. Such collaboration culminates in cultivating a more robust, efficient, and adaptable supply chain ecosystem.

The rise of the OR-AI synergy in the business domain

Going back to Jack Levis (the initial subject of our story), he eventually became a driving force within his company and beyond by leading what is now considered the largest analytics effort worldwide, which resulted in a platform that analyzes data to optimize UPS drivers’ routes. On-Road Integrated Optimization and Navigation (ORION) every year saves UPS 10 million gallons of fuel, 100,000 metric tons of greenhouse gas emissions, 100 million driven miles, and $300-$400 million dollars. ORION examines more than 250 million distinctive address points alongside clients’ shipping specifications (such as time constraints) to determine the best route, given the presence of multiple conflicting factors along the way (such as traffic and weather patterns which are analyzed through ML). These routes, in some cases, may sound counterintuitive to humans, but for a machine that can process an immense number of data points, it makes perfect sense. Jack Levis later became Senior Director of Process Management for UPS and, finally, shared his story of perseverance in the face of setbacks in this TED talk.

Just like UPS, many other companies have reported very successful synergies between OR and AI techniques. For instance, since its origins several decades ago, INFORMS (Institute for Operations Research and the Management Sciences) has organized the Franz Edelman Award for Achievement in Advanced Analytics, Operations Research, and Management Science, the world’s most prestigious award for achievement in the practice of analytics and operations research (sometimes called the “Nobel”). Last year, the winner was the Government of Chile for its successful partnership with the Instituto Sistemas Complejos de Ingeniería and the telecom company Entel in the use of methodologies and tools that placed OR and ML tools at the forefront of the battle against the COVID pandemic. This year, Walmart’s Global Technology team won with an application supporting daily decisions of truck routing and loading using advanced optimization models to build a transformation roadmap for long-term supply chain capital investments.  The application avoided 72 million pounds of CO2 and resulted in savings of $75 million during fiscal year 2023. Here is a list of some of the other finalist companies and the types of impact they have reported:

How Fractal Does It

By combining the power of AI and OR, we have helped businesses optimize their operations, improve customer satisfaction, and increase revenues. Our solutions have been used to identify new revenue streams, optimize supply chains, and reduce costs. These are examples of our success:

Case 1: Truck Build Project

Fractal helped a client generate $3 million in additional weekly shipments by automating its truck-building process through a multi-objective lot sizing and load optimization algorithm. The algorithm incorporated order priorities and business constraints, balancing revenue, profit margin, and service levels. Fractal provided a user interface that allowed to modify constraints and monitor load build metrics over time, resulting in a more efficient and optimized process.

Case 2: Packaging

In another project, Fractal improved packaging efficiency by creating optimized packaging schedules that considered available capacity and reduced the demand-supply gap while minimizing pack-size changeovers and processing time. The solution decreased average change-over time by 5%, achieved an average demand-supply gap close to 2.5% versus the current state of an 8% gap, and identified potential risks of over and underutilization of machines.

Case 3: eComm Fulfilment

Fractal helped a retail client decrease delivery time to 2 days instead of 4-5 days for 90% + of shipments. The solution included maximizing speed to customers with minimum fulfillment cost, optimal inventory placements and mix, and reducing the number of splits during fulfillment. By implementing this optimization solution, the client could provide faster and more efficient delivery services to their customers.

Case 4: Inventory shipping

Fractal helped a retail client replace its existing mainframe inventory shipping solution with a modern, efficient, and maintainable optimization algorithm. This algorithm decreased sales by more than 1500% and order fulfillment efficiency by 2%, while respecting store pick priorities, prioritizing allocations with high risk to the service level, deciding when to over/under allocate, and leveraging proportional allocation targets to ensure allocational fairness when considering different pack types. As a result, the client achieved an optimal allocation of resources, resulting in a more efficient and cost-effective process.

Conclusion

At Fractal, innovation is not just a buzzword; it’s our driving force. Our commitment to staying at the forefront of technological advancements means continuously refining our AI and OR methodologies. As a result, our clients benefit from the latest breakthroughs, always one step ahead of the competition.

Our team of experts comprises leading data scientists, AI specialists, and seasoned operations researchers, working in tandem to create tailor-made solutions for each client. By blending the powers of AI and OR, we provide an unparalleled competitive advantage, solving complex problems with robust acumen and visionary foresight.

Imagine a world where your business runs like a well-oiled machine, automating routine tasks, anticipating market changes, and fortifying every facet of your operations. Fractal transforms this vision into reality, leading your company to unprecedented success. Join us on this transformative journey, where the perfect blend of AI and OR will redefine what’s possible for your business. Embrace innovation, seize opportunities, and supercharge your growth with Fractal today!

BIBLIOGRAPHY

Bar-Yam, Y. & Bialik, M., 2013. Beyond Big Data: Identifying Important Information for Real World Challenges. [Online] Available at: https://necsi.edu/beyond-big-data-identifying-important-information-for-real-world-challenges

Davenport, T. H. & Harris, J. G., 2017. Competing on Analytics: The New Science of Winning; With a New Introduction. 2nd. ed. Boston, Massachusetts: Harvard Business Review Press.

Dawson, C. S., Murphy, B., Wolman, E. & McCallum, Jr., C. J., 2020. Operations Research at Bell Laboratories through the 1970s: Part I. Operations Research, 48(2), pp. 205-215.

Gass , S. I. & Assad, A. A., 2011. History of Operations Research. Transforming Research into Action, Volume September, pp. 1-14.

Goodeve, C., 1948. Operational Research. Nature, 161(4089), p. 377–384.

INFORMS, 2023. Operations-Research-Analytics. [Online] Available at: https://www.informs.org/Explore/Operations-Research-Analytics

INFORMS, n.d. @INFORMSonline. [Online] Available at: https://www.youtube.com/@INFORMSonline

INFORMS, n.d. Recognizing Excellence. [Online] Available at: https://www.informs.org/Recognizing-Excellence

Keeman, P. T., Oven, J. H. & Schumacher, K., 2022. Introduction to Analytics. In: Analytics Body of Knowledge. s.l.: INFORMS, pp. 1-21.

Ravindran, A. R., 2009. Operations Research Methodologies. Boca Raton, FL: CRC Press. Shrader, C. R., 2006. History of Operations Research in the U.S. Army. Volume I: 1942-1962.

Washington, D.C.: Office of the Deputy Under Secretary of the Army for Operations Research – United States Army.

Subramanian , A. & Teichgraeber, H., 2023. The Interplay Between Operations Research and Machine Learning. ORMS Today, 02 02.