Extending Kirstine

Since Kirstine.jl's code base is quite modular, you can extend the package's functionality in multiple ways. All of the high-level logic is implemented in terms of abstract supertypes. For example, a general DesignRegion is used wherever possible, and the specifics of a DesignInterval are implemented in a few short functions. Hence you can, in your own code outside the package, define a custom subtype of DesignRegion and the handful of required methods, and it will seamlessly blend in. Such is the magic of multiple dispatch. This does not only work with the parts of a DesignProblem, but also with the optimization algorithms. They are represented by subtypes of Kirstine.Optimizer (for the low-level work) and Kirstine.ProblemSolvingStrategy (for high-level design-specific stuff). Moreover, the finicky logic for moving around DesignMeasures is encapsulated, so your custom Kirstine.Optimizer subtype will only have to deal with AbstractPoints. The following example vignettes will show you how to implement

• a non-logarithmized determinant design criterion
• an ellipsoidal design region
• a nonlinear logistic regression model
• a regularizing normal approximation
• a variant of the imperialist competitive algorithm as a low-level optimizer
• a multi-run problem solving strategy