Givaudan Works With MIT on Flavor Algorithm

In a collaboration with the Massachusetts Institute of Technology (MIT), Givaudan is working on using “genetic programming” to create an algorithm that helps better investigate and discover consumer taste preferences.

In an article from the MIT News Office by Larry Hardesty titled “The mathematics of taste,” the researchers explain the necessity and difficulties of consumer testing for flavor products. The article notes that, “Subjects’ preferences can vary so widely that no clear consensus may emerge” and can also suffer from “smell fatigue” after number of samples. To combat these problems, Givaudan and members of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) are using genetic programming via mathematical models to determine preferences.

The mathematic models work to fit the available data of taste test results, which then work to “cross-pollinate” for even more accurate representative models. Givaudan worked with the CSAIL researchers Una-May O’Reilly, Kalyan Veeramachaneni and Ekaterina Vladislavleva, providing data for the researchers to interpret the results of data from 69 test subjects who tasted seven basic flavors in 36 different combinations by randomly generating mathematic functions that would predict taste test scores based on the concentrations of the different flavors in the combinations. According to the article, “Each function is assessed according to two criteria: accuracy and simplicity. A function that, for example, predicts a subject’s preferences fairly accurately using a single factor—say, concentration of butter—could prove more useful than one that yields a slightly more accurate prediction but requires a complicated mathematical manipulation of all seven variables.” The CSAIL researchers then dispensed with the functions that produced poor outcomes, and various elements of the remaining functions were randomly combined to create a new batch of functions, which are again evaluated for accuracy and simplicity. This process is repeated approximately 30 times until a set of functions that mirrors the preferences of a single test subject well emerges.

As the method also produces profiles of individual test subjects’ tastes, it can sort them into groups, which can then be factored into flavor creation and product marketing. The CSAIL researchers also developed a set of functions to test the accuracy of their models—creating first functions that represent a subject’s true taste preferences and then showing how, given the limitation of particular test designs, the algorithms could still discover those preferences. This creation has garnered interest from flavor researchers who wish to use it to develop better flavor and taste test protocols.

More in Regulatory & Research