A Culinary Computational Creativity System - Computational Creativity Research: Towards Creative Machines - page 326

further constrain the combinations: ingredients that either do not pair well with the
base ingredient, or are not found in the target cuisine, are eliminated from the pos-
sibilities.
Additional filters could be applied. Most dietary restrictions (vegetarian, low-fat,
gluten-free, Kosher) can be mapped to a list of ingredients to avoid, sometimes by
checking the ingredient types or their nutrition facts. In another example, a profile,
either built explicitly by the users or inferred from their online history, may indicate
preferred foods and help produce personalized recipes.
In most cases, the above approaches are not sufficient to narrow down the com-
binations to a number that can be handled by today's computers in real time. The
work product assessor can alleviate this issue by retaining only the recipes that rank
the highest with respect to one or more evaluations, such as novelty or estimated
pleasantness. Ultimately, if too many combinations subsist, the recipes displayed to
the user may be chosen randomly. Generating random subsets of possibilities has the
added benefit of ensuring that every run of the system, even with the same inputs, will
produce slightly different results, thus avoiding repetition and keeping the system
creative.
Recall that the work product assessor ranks ideas according to metrics defined
using basic ideas from information theory, psychology, and chemistry to predict the
novelty and flavorfulness of newly-created ingredient combinations. This approach
does not require training data on full recipes/dishes and is distinct from a super-
vised learning approach [ 14 ]; the benefit is that the system is more likely to create
novel and untrained ingredient combinations. Data sources from culinary traditions,
chemoinformatics, and hedonic psychophysics are, however, required.
16.4 Calculating Ingredient Proportions
Once the list of ingredients for the new recipe has been generated by the system,
the next step is to calculate the ingredient proportions in the new recipe. Since our
system is almost entirely data-driven, striving to generate a creative recipe using
a corpus of recipes, we use the corpus itself to help guide these calculations. Our
goal is to try to match the proportions in the new recipe with certain distributional
characteristics of existing recipes, for instance, the distribution of ingredients and of
certain nutritional coefficients. This bears strong resemblance to methods for texture
synthesis [ 22 , 23 ].
When measuring ingredients for a new recipe, chefs rely on several principles,
depending on the nature of the dish being created. Desserts and pastries (sweet and
savory) require precise measurements of ingredients to achieve a nutrient composi-
tion that in turn creates a specific texture. In the example of a custard, the amount
of egg protein (and its nature, depending on whether one uses yolks or whites) as
compared to the amount of liquid determines the firmness of the cooked mixture.
Minerals are also required, so as to obtain a coherent gel instead of curdled eggs
floating in liquid [ 24 , p. 94]. In cake recipes, the impact of the proportions of various