Model-Based Automated Generation of User Interfaces

Angel R. Puerta, Henrik Eriksson, John H. Gennari, Mark A. MusenAAAI 1994

Introduction

In the early 1990s, user interface (UI) development represented a significant bottleneck in software engineering. While early "interface builders" could manage static layouts—positioning buttons and text fields—the dynamic behavior of an application remained a manual, code-intensive task. This includes window navigation, conditional field visibility, and data flow logic. This paper introduces Mecano, a system developed by the Stanford Medical Computer Science Group to automate the generation of both static and dynamic UI components directly from high-level domain models.

The Mecano Architecture

The core philosophy of Mecano is that standard data models lack the semantic richness required for full UI automation. Instead, it utilizes domain models: knowledge representations that encapsulate the rules, relationships, and metadata of a specific field.

The Mecano system architecture, illustrating the relationship between design tools, models, and the run-time system.

The development process follows a four-stage pipeline:

  1. Domain Model Specification: Developers define classes and their relationships (e.g., "is-a" hierarchies and "part-of" groupings) using a specialized Model Editor.
  2. The Intelligent Designer: This component analyzes the domain model and applies heuristic design rules. For instance, if a data field is marked with "multiple cardinality," the system automatically selects a list-browser widget.
  3. Interface Model: The output is a platform-independent declarative model describing every window and interaction rule.
  4. Run-Time System: This layer interprets the interface model to render the final UI on the target platform (originally NeXTSTEP).

Driving UI Behavior through Facets

A key innovation in Mecano is the use of "facets"—metadata attached to data slots that inform the UI generation process. By defining properties such as data types or allowed classes, the system can make intelligent decisions about widget selection and validation logic.

An example of a domain model for chemotherapy, showing slots and their associated facets which drive UI generation.

As shown in the figure above, a "Chemotherapy" class might have a slot for "Drug_Parts" with a facet restricting it to the "drug" class. The Intelligent Designer uses this information to automatically generate a specialized selection interface or navigation link to drug-related windows. This coupling ensures that if the underlying medical protocol changes in the model, the UI updates to reflect it automatically.

Evaluation: Medical Therapy Administration

The system was evaluated using the highly complex domain of chemotherapy protocols. These protocols involve hundreds of data points and strict procedural workflows. Mecano successfully generated a massive medical interface consisting of over 60 windows and hundreds of widgets.

Remarkably, the manual "custom-tailoring" phase—where developers tweak the automated output to meet specific user preferences—required only 2.5 to 4 hours for the entire 60-window application. In traditional development environments of the time, building an interface of this complexity would have taken weeks or months of engineering time.

Conclusion and Future Work

Mecano demonstrates that model-based approaches can significantly reduce the effort required to build complex, data-driven interfaces. By leveraging the inherent structure of domain knowledge, navigation and widget selection can be largely automated. However, the system remains dependent on the quality of the underlying model. The authors acknowledge that future iterations should incorporate "task models" to better capture user intent and further refine the generated interaction sequences.

Summary updated

BibTeX
@inproceedings{puerta1994model,
  author = {Angel R. Puerta and Henrik Eriksson and John H. Gennari and Mark A. Musen},
  title = {{Model-Based Automated Generation of User Interfaces}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure