AI model solves optimization problem
Technically speaking, this is a classic optimization problem: What must a coffee manufacturer do to create a specific flavor profile at a specific price? DTI modeled this problem by way of a prototype and a neural network. Different data such as bean types, blends, and tastes formed the basis here, although the difficulty was that neural networks normally learn from a huge amount of data and from many past examples. In this case, however, the database was very small.
In their audit, the IPA experts therefore recommended dispensing with the neural network and instead using a model with a simpler structure and fewer parameters. Ultimately, a new model featuring fewer than ten parameters was used.
In addition, the IPA experts divided the optimization problem into two smaller sub problems:
- What is the optimal blend for a specific flavor profile, given a certain pre-selection of beans?
- Which beans are best-suited for this, also in terms of price?
DTI proposed using what is known as a genetic algorithm, which can solve combinatorial problems by “trial and error”, with other algorithms also used to supplement this.
The simplified model also has the advantage of acting more predictably than neural networks. The latter are commonly referred to as “black boxes” because even experts are often unaware as to how exactly a result was actually obtained. This lack of explainability can reduce trust in an AI application and is often a legal hurdle as well.