Applicable Instances is an R package designed to fine tune instance generators. The tools implemented in this package enable users to change a generator so that the instances produced are applicable to the conclusions of an experiment, i.e., the instances:
- help highlight differences in performance between algorithms, and
- are similar in a formal way to a seed set of instances, presumed to be real instances.
When the instance generator fails to produce instances with the characteristics above, this package can pinpoint exactly what needs to be improved.
The package is focused on Branch Problems, i.e., problems with a complex structure, typically involving a combination of stereotypical problems (Core problems like TSP or QAP) and additional side constraints. The difficulty of any given instance of a Branch Problem is typically strongly dependent on the specific data that defines that instance. Thus the need to make sure that generators produce instances that are applicable to the conclusions sought, whether those conclusions are academic in nature or they are a motivated by a comparison between competing software packages.
The details are in the paper:
Leo Lopes and Kate Smith-Miles, Generating Applicable Synthetic Instances for Branch Problems,
The core methodological tool used in the package is the Classification Tree (sometimes called a Decision Tree, but not to be confused with the homonym tool from Decision Analysis). The latest technical documentation of the package is available here: