Learning Design Guidelines by Theory Refinement

Jacob EisensteinAAAI 1999 · OpenReview

Introduction

The automation of user interface (UI) design is fundamentally a problem of generating mappings between abstract specifications and concrete interface components. While systems like the MOBI-D environment attempt to formalize these mappings, they often encounter a significant hurdle: the "intuition" of human designers. Individual designers frequently follow stylistic preferences or "schools of design" that are not easily captured in static formalisms. When automated tools ignore these nuances, they forfeit the flexibility and creativity inherent in human design.

To address this, we propose an adaptive approach to automatic design. By integrating theory-refinement algorithms into the design process, automation can proceed without an exhaustive initial knowledge base, instead refining its logic through interaction. This allows the system to accommodate stylistic preferences and adapt to changing technological requirements.

The TIMM Module and Decision Trees

This research focuses on the TIMM module within the MOBI-D user-interface design environment. MOBI-D maintains explicit, formal representations of both abstract domain objects and concrete presentation elements. TIMM automates the mappings between these two sides using a decision tree.

While many theory-refinement systems utilize neural networks, decision trees offer a more transparent structure for UI design logic. When a designer disagrees with a choice made by TIMM, the system treats this as an error and initiates a refinement process to align its internal logic with the user's intent.

Adaptation via Theory Refinement

The system employs a greedy hill-climb search to identify a set of operations that minimizes errors across the user's entire interaction history. To maintain model stability and prevent over-correction, the algorithm follows a strict hierarchy of preferred operations based on their reversibility and structural impact:

  1. Leaf Output Change (Most Preferred): The system alters the output of the specific leaf that caused the error. For example, if the tree suggests a checkbox for a boolean variable but the user prefers radio buttons, the leaf's output is updated. This is the most preferred operation as it is easily reversible and does not alter the tree's structure.
  2. Boundary Condition Alteration: If a leaf change is insufficient, the system attempts to shift the thresholds of existing discriminants. For instance, if a list box is triggered at four items but the user prefers it only for five or more, the boundary is adjusted. This remains preferable to structural changes as it can be reversed by further boundary shifts.
  3. Adding a New Branch (Least Preferred): When an error stems from a relevant piece of information the tree currently ignores—such as the size of a parent window—the system adds a new discriminant. This is considered the least preferred operation because it is irreversible within the current framework and fundamentally changes the tree's structure.

Insights and Results

Preliminary testing indicates that these operations effectively bridge the "intuition gap." What is often perceived as "intuition" is frequently just a set of relevant variables that have not yet been formalized. By allowing the system to discover these variables (like window size or stylistic trends) over time, the knowledge base evolves alongside the designer's needs.

Key findings from initial evaluations include:

  • Search Efficiency: Despite the use of a greedy hill-climb search, the system did not encounter local minima during testing, suggesting the search space is well-behaved for this domain.
  • Versatility: The algorithm demonstrated success in acquiring new design knowledge, accommodating idiosyncratic user preferences, and updating automation logic to reflect changes in technology.
  • Reversibility as a Constraint: Prioritizing reversible changes helps prevent the system from "overfitting" to temporary user whims, ensuring the core design theory remains robust.

Limitations and Future Directions

While the current system successfully adapts to user feedback, it lacks a mechanism for removing branches. This irreversibility means that over long periods of use, the decision tree could become unnecessarily bloated with discriminants that are no longer relevant. Future work should explore methods for pruning or retiring branches. Additionally, while preliminary results are promising, more extensive testing is required to determine how the search complexity scales with significantly larger interaction histories and more diverse design domains.

Summary updated

Abstract

The automation of design decisions can be seen as a problem of generating mappings between elements in an abstract specification of the object to be designed and the concrete parts of the object itself (Puerta and Eisenstein 1999). In some cases, it is difficult to discover a formalism that takes all relevant variables into account; human designers proceed by Ointuition.O Individual designers may have stylistic preferences that are purely idiosyncratic or are common only to one particular Oschool.O By ignoring such preferences, automatic design forfeits the flexibility, creativity, and vitality of human design. In short, automatic design algorithms suffer from a lack of flexibility. Adaptation is offered as a solution to this problem. By making automatic design algorithms adaptive, we can begin to do automation without a complete knowledge baseNit can be developed and refined along the way. Stylistic preferences can likewise be accommodated if an automatic design algorithm can adapt to the user. We add adaptation to an existing piece of automatic design software: the TIMM module of the MOBI-D user- interface design environment (Puerta 1997). MOBI-D maintains explicit, formal representations of the abstract and concrete sides of the interface. TIMM automates the mappings between the abstract domain objects and concrete presentation elements, using a decision tree. Although there exists a body of work on using decision trees to automate selection of interactors (Vanderdonckt and Bodart 1996), most theory-refinement algorithms involve neural nets (Maclin and Shavlik 1996). A theory- refinement algorithm for decision trees is presented below. An error occurs when the user disagrees with one of TIMMOs design decisions. The system then searches for a set of operations that can be performed on the decision tree so as to yield the least number of errors, when applied over the entire history of user interactions. When multiple solutions produce the same number of errors, operations that are reversible and do not affect the structure of the decision tree are preferred. The most preferred operation is to change the output of the leaf that produced the error. Suppose that the decision tree dictates that boolean domain elements are best treated by a checkbox, but the user selects radio buttons instead. This stylistic preference can be satisfied by changing the output of the leaf for boolean domain variables to radio buttons. An alteration of the boundary conditions is the next most preferred operation, because it can be reversed by altering the boundary conditions again. Suppose that the decision- tree selects radio buttons for variables with less than four allowed values, and list boxes when there are four or more allowed values. If the user rejects the selection of list boxes in cases where there are four allowed values, then a shift in the boundary condition will produce more agreeable output. There is no facility for removing branches. Thus, the addition of a new branch is the least preferred operation, as it is irreversible. It is necessary to add a branch when there is a relevant piece of information that the decision tree did not consider. The designerOs selection of certain interactors might be influenced by the size of their parent dialog window. Adding a new discriminant to take that information into account is the only way to properly correct the decision tree. Preliminary testing has shown that all of these operations are useful. Local minima have not appeared, so the fact that our search is a greedy hill-climb does not pose a problem. Earlier, three advantages of adaptation were cited: acquisition of new design knowledge, accommodation of user preferences, and update of the automation algorithm in response to changing technology. Testing has shown that the theory-refinement algorithm described here can deliver all three of those advantages.
BibTeX
@inproceedings{eisenstein1999learning,
  author = {Jacob Eisenstein},
  title = {{Learning Design Guidelines by Theory Refinement}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1999},
  url = {https://openreview.net/forum?id=rkZndRlubB},
}

Loading related content...

Figure