A Culinary Computational Creativity System - Computational Creativity Research: Towards Creative Machines - page 333

different stages without compromising the result; oil is generally used for frying,
and its position in the graph is mainly determined by which other ingredients need
to be fried. A distinction between primary (such as proteins) and secondary (such
as seasoning) ingredients emerges as a useful classification that can be added to the
knowledge database.
The final graph produces high-level recipe instructions. By looking back at the
ingredient subgraph clusters that were retained, it is possible to make a second pass
to refine the specifics of each task, from the oven temperature to how vegetables
should be chopped.

16.6 Estimating Recipe Step Durations

To enable work planning, it is important to estimate how long a given recipe will take
to prepare. Since we are constructing completely new recipes, we must estimate the
durations of individual recipe steps and then sum them together to find an estimate of
the complete recipe's duration. Unfortunately, durationmeasurement data is typically
available for complete recipes rather than for individual steps, e.g. we may have data
on how long it takes to prepare a crostini, but not how long it takes to peel and cube
lychees. This section describes an approach for “unmixing” durations of individual
steps from data pertaining to complete recipes.
Before proceeding, we provide the following definitions:

￿ Atom : a work task that cannot be broken into smaller constituents,

￿ Molecule : a structured collection of atomic tasks that are linked together as a larger

task, e.g. through a directed acyclic graph,

￿ Equivalence Class of Atoms : a set of atomic tasks that are thought to require the

same amount of time,

￿ Molecule Catalog : a predefined list of all possible molecules, their constituent

atoms, and their durations,

￿ Incomplete Atom Catalog : a predefined list of all possible atoms and which mole-

cules contain them, and

￿ Complete Atom Catalog : a predefined list of all possible atoms and their durations.

The core idea is that if we have estimates of the duration of a large number of
complete recipes (molecules), with overlapping sets of individual steps (atoms), we
can use inference algorithms to find the effort of the individual steps by solving an
inverse problem. The basic approach is depicted diagrammatically in Fig. 16.10 and
proceeds as follows:

1. Each molecule in a catalog is broken down into its constituent atoms,
2. The resulting incomplete atom catalog is categorized into equivalence classes,
3. A measurement operator is constructed from the listing of which atom is part of
which molecule, and a measurement vector is constructed from the durations of
the molecules,