Design and Implementation of a Two-Phase AI System to Predict Quality Features and Optimize Recipes of Baking Goods

  • Manuel Gabriel Wechselberger

Student thesis: Master's Thesis

Abstract

Nutritional considerations are becoming increasingly important in the baking sector,
as well as in other food-related sectors. The development of products meeting specific
nutritional requirements is a time-consuming process, as the complex interactions between ingredient composition and quality characteristics often require a trial-and-error
approach. This master’s thesis investigates the possibility of supporting the development
process with a well-designed artificial intelligence (AI) system. The proposed AI system makes suggestions on how to alter recipes to achieve target nutritional values while
preserving product quality. Therefore, predictive modelling and optimization techniques
are utilized in a two-phase approach. To ensure product quality is preserved, predictive
models that predict the area of a product and a change in the quality feature crispiness are integrated as constraints during the optimization process. The most accurate
model for predicting changes in crispiness was a random forest model, which achieved
an area under the receiver operating characteristic (AUROC) curve score of 0.940. This
model was trained on carefully generated synthetic data that simulates the real-world
development process. The most accurate model for predicting the area of a product was
a LightGBM model, achieving an AUROC score of 0.961, and was trained on actual
product data. By utilizing the covariance matrix adaptation evolution strategy (CMAES), nutritional values for 68 out of 201 recipes were successfully optimized under the
restriction of the constraints. Furthermore, a reinforcement learning (RL) approach was
investigated, using one-shot RL and an actor-critic method for continuous optimization.
However, this approach was unable to optimize recipes within the constraints. Despite
this, the AI system shows significant promise, although further testing with real-world
data is necessary to fully evaluate its potential and applicability.
Date of Award2024
Original languageEnglish (American)
SupervisorStephan Winkler (Supervisor)

Cite this

'