Automatically generating personalized user interfaces with Supple
Interface generation as optimization
Download 5.78 Mb. View original pdf
|
| 4. Interface generation as optimization The goal is to render each interface element with a concrete widget, as illustrated earlier in Fig. 1. Thus a legal rendering of a functional specification S f is defined to be a mapping R : S f → W which satisfies the interface and device constraints in C I and C D . Of course, there maybe many legal renderings. Therefore, in order to find the best one, Supple relies on a cost function $ : R → R, which provides a quantitative metric of the user interface quality. The cost function can correspond to any measure of quality of a user interface, such as consistency with the user’s stated preferences (Section 5.1) or expected speed of use (Section 5.2). It can also incorporate additional concerns, such as similarity to previously seen renderings of a user interface, even if those renderings were generated for other devices (Section We thus formally define the interface rendering problem as a tuple I, D, Ti, where I ≡ S f , C I abstractly describes the interface in terms of the functional specification and the interface constraints, D ≡ W, C D is a device model specifying available widgets and device constraints, T is the usage trace, and $ is the cost function. R is a solution to a rendering problem if R is a legal rendering with minimum cost—we thus cast interface generation as constrained optimization, where the goal is to find a concrete user interface that minimizes the expected value of the cost function with respect to the usage trace, subject to the interface and device constraints. As stated, this is a hard discrete/continuous hybrid problem because W contains different classes of widgets, each of which is parametrized with several real parameters, such as minimum target size s t , minimum visual cue size s c , and additional widget-specific parameters, for example, the length of a list widget for showing search results in the Amazon search interface (Fig. 4), can vary reasonably from a handful up to 40 entries. K.Z. Gajos et al. / Artificial Intelligence 174 (2010) 910–950 919 Table 1 An algorithm combining branch-and-bound discrete optimization and constraint satisfaction mechanisms. The variables correspond to the elements in the functional specification S f , their possible values are drawn from the set of available widgets W , and the constraints include both interface and device constraints (i.e., C I and C D ). The solution is stored in bestRendition. bestCost ← ∞ bestRendition ← null Download 5.78 Mb. Share with your friends:
|