Content area
(Human tasters still required.)" * "A new tool developed by Mondelez 15 speeding up creation of snack recipes and optimizing them to fit certain taste profiles." * Marion Nestle, professor emerita of nutrition, food studies and public health at NYU, said that "Food companies like Mondelez are racing to try out AI in every area of their business, from supply chains to marketing to recipe development." * Its food scientists use the AI tool to create optimal recipes according to desired characteristics such as flavor, aroma and appearance, as well as ingredient costs, environmental impact and nutritional profile. If all the equations are linear per LP (linear programming), the Simplex Method developed by George Dantzig around 1947 can be used to solve for optimal decisions (with a calculator in his day). In my class, 1 teach that back in the day (around the 1960s), the Simplex Method was being used by various corporate functions including advertising, finance, and operations/manufacturing. Figure 1 is a copy of Royal Burger's optimized Solver spreadsheet using the Simplex Method. Figure 3 is a copy of the AppleSmash's optimized Solver spreadsheet using the Simplex Method.
Yes, optimization is a core part of AI, often requiring powerful computing to solve complex problems, and it raises the question of whether future Al systems could one day surpass human-designed algorithms with their own.
A year ago, I wrote an Insights column titled: "AI update: Decision-maker or decision-supporter?" (July/August 2024), because artificial intelligence was once again being hyped by the tech community. I've been keeping track of AI since the early 1970s when I read a book, "The Sciences of the Artificial" by Herbert A. Simon, a professor at Carnegie Mellon. It discussed learning sciences and artificial intelligence. I have followed IBM's progress in AI with its development of a chess-playing computer system known as Deep Blue and Watson (a questionanswering computer system capable of responding in natural language). To date, however, IBM has been less than successful when it comes to significantly selling AI products and services. In addition, I haven't seen any "Blockbuster" Al applications generating significant revenues for tech providers. The current activity is largely supply-driven, with tech competitors focused on creating cloud-based platforms for AI app developers on the mantra of "build them, and they (the developers) will come." A lot of money is being invested in large-scale, cloud-based server farms, along with the electricity to run them.
Simply put, Al is largely about getting a computer to think better than humans, often based on replicating our brains" neural network structure. I'm specifically interested in whether AI can help us make better business decisions toward bettering economic prosperity and everyday life. The current hype in AI started with the launch of ChatGPT. As I wrote, it is very impressive in this regard. However, the tagline to the column was: "The promise of Al offers greater decision-making power, but there are some decisions that Al should not make." In effect, Al needs to support decisions not necessarily make them for us-its decisions need humane, human supervision.
Mondelez is tapping AI
There was an interesting online version of an article in the Wall Street Journal (WSJ) dated Dec. 17, 2024 and titled: "Oreo Owner Mondelez Taps AI to Tweak Its Classic Snacks." The article was intriguing, and I took away from it a few salient points.
* The tagline stated: "The cookie- and candy-maker's R&D lab has gone into overdrive with AI, which is updating some of the company's treats and spinning out new iterations. (Human tasters still required.)"
* "A new tool developed by Mondelez 15 speeding up creation of snack recipes and optimizing them to fit certain taste profiles."
* Marion Nestle, professor emerita of nutrition, food studies and public health at NYU, said that "Food companies like Mondelez are racing to try out AI in every area of their business, from supply chains to marketing to recipe development."
* Its food scientists use the AI tool to create optimal recipes according to desired characteristics such as flavor, aroma and appearance, as well as ingredient costs, environmental impact and nutritional profile.
I was initially intrigued by the article because it dealt with using AI to optimize food products, the first I've seen of this.
What is optimization?
During my retirement, I teach one course in the business school at the Lowell campus of the University of Massachusetts. The undergraduate course is titled: "Introduction to Business Analytics." There are three categories of business analytics: 1) descriptive, 2) predictive and 3) prescriptive. Optimization algorithms are "prescriptive" in that they tell decision-makers what decisions need to be made to optimize (i.e., maximize or minimize) an objective function while satisfying a set of constraints (e.g., maximum and minimum limits). Constraints include an organizations supply, as well as customer demand (for example).
Of course, the optimal decisions are based on a quantitative model that precisely represents the real world. If all the equations are linear per LP (linear programming), the Simplex Method developed by George Dantzig around 1947 can be used to solve for optimal decisions (with a calculator in his day).
I discussed constrained-based optimization in the Insights column: "The special demand optimization team" (September/October 2024) With regard to optimal supply-demand planning under severe supply shortages. In my class, 1 teach that back in the day (around the 1960s), the Simplex Method was being used by various corporate functions including advertising, finance, and operations/manufacturing.
There is also a class of LP problems involving "blending" formulations, wherein one 15 looking for the lowest cost mix of various ingredients. One I discuss in class is a food-diet problem for a turkey farm that raises the birds for holiday sales. The farm wants to blend two brands of feed to help the turkeys meet or exceed their "minimum monthly ingredient requirements." However, I always felt that it was not interesting enough for students because turkeys don't care about the feed's taste. They are not picky, and will eat what you feed them.
Luckily, several years ago, I encountered an article in the WSJ (June 5, 2019) titled: "Meatless Burgers Add Some Sizzle and Demand at Fast-Food Chains." It talked about Beyond Meat and Impossible Foods developing meat-free beef and marketing it to fast-food restaurants. At the same time, the Dunkin Donuts outlet on campus was launching a meat-free "Beyond Sausage" sandwich. So, I decided to put a problem on my next exam about a fictitious fast-food chain named Royal Burger. The chain was planning to blend three brands of fake beef, to achieve the characteristics needed for their meat-free burgers at the lowest cost.
In addition, back in my consulting days, I was given a tour of a Motts plant that crushed and blended various types of apples to make applesauce. The sauce was made to align with their TV ad jingle: "The finest apples from Apple Land/Make Mott's Apple Sauce taste grand!" The apple sauce also had to taste the same each time. I decided to make that an exam question, as well. The fictitious company's name was AppleSmash, which "was optimally blending three types of apples, to meet requirements at the lowest cost."
While reading the WSJ article about Mondelez it struck me as uncanny that these exam questions are remarkably similar to what the company was striving to do. The two exam questions are described below.
The Royal Burger question
The Royal Burger fast-food chain is planning to add meatless burgers to its menu. It has been testing fake beef from three suppliers: NotLikely, AfterMeat, and Delicious. Delicious is a startup with a very flavorful fake beef that's inexpensive, but has high salt content. Royal is using a Blended LP model to determine what blend of fake beef gives the lowest cost per pound. The model solves for the weight of each fake beef in 16 ounces of the blend. There are 3 constraints on the blend: 1) salt content in milligrams (mg) must be 3,000 or less; 2) a composite Flavorful Score-in which each supplier's fake beef has been rated from 1 to 10- which must be 120 or more out of a possible highest score of 160 (=10·16 07); and 3) protein content must be 200 mg or more.
Figure 1 is a copy of Royal Burger's optimized Solver spreadsheet using the Simplex Method. (Solver is an add-in to Microsoft's Excel product.) The cost of the optimal mix is $2.70 per pound. That represents 6 ounces of NotLikely, no AfterMeat, and 10 ounces of Delicious. The RHS (right-hand side) of the constraint equations-the maximum and minimum limits-and the LHS (left-hand side) of the constraint equations are shown in the bottom right of Figure 1. Namely the optimal mix's: salt content, flavorful score, protein content, and weight, respectively.
The AppleSmash question
AppleSmash crushes three (3) types of apples to make applesauce that 1s "never too bitter, yet tart and sweet as can be." Operations processes 1,000 or more barrels of apples each month by crushing and blending three (3) types of apples: Cortland, Macs, and Jazz. The operating and apple costs, plus flavor content for each type of apple per barrel are shown in the table in Figure 2. In addition, marketing limits the flavor content of processed barrels to be 24,000 or less for bitterness, at least 15,000 for tartness, and at least 20,000 for sweetness. Manufacturing has only 200 barrels of Jazz apples available at the start of a month.
Figure 3 is a copy of the AppleSmash's optimized Solver spreadsheet using the Simplex Method. The lowest cost of the optimal mix is $40,000, representing no Cortland, 900 barrels of Macs, and 100 barrels of Jazz, respectively. The RHS of the constraint equations- the maximum and minimum limits-and the LHS of the constraint equations are shown in the bottom right. Namely the optimal mixes: barrels of Jazz crushed, the contents of flavor for bitterness, tartness, and sweetness, and the barrels of apples crushed in total, respectively.
Closing comments
These optimization formations might be useful for food companies such as Mondelez. The questions of food characteristics include not only food recipes, but also quantities of taste (such as bitterness, sweetness, and tartness) and flavorful scores. Can companies develop quantitative metrics for these? Can AI machine-learning systems find them on the worldwide web? Probably not, and that is why the WSJ article stated that: "Human tasters still required." I suspect that will be true.
Is optimization part of AI? I say yes, because one's human brain is not powerful enough to solve large-scale, complex optimization models. Powerful computers are needed to solve realistic optimizations as quickly as possible, even in real-time. Will AI systems always require man-made algorithms? Or, someday, will it be possible that AI "god-like models (such as ChatGPT) will develop faster algorithms than we mortal humans can? ·
Copyright Peerless Media 2025