Introduction
In the field of Human-Computer Interaction (HCI), cognitive models serve as essential tools for simulating human behavior and predicting how users interact with software interfaces. However, researchers often face a significant trade-off between model simplicity and fidelity. High-level frameworks like GOMS (Goals, Operators, Methods, and Selection Rules) are relatively straightforward to construct but offer only coarse timing estimates. Conversely, detailed architectures like ACT-R (Atomic Components of Thought-Rational) provide high-fidelity simulations of memory, visual attention, and motor control, but are notoriously difficult and time-consuming to program manually.
The construction of an ACT-R model for even a simple task can take weeks of expert labor. To address this "modeling bottleneck," the G2A system was developed to automate the translation of abstract GOMS models into detailed ACT-R architectures, reducing the development time from weeks to hours.
The G2A System: Translation as Search
The core innovation of G2A is treating the translation process not as a simple compilation task, but as an AI search problem. Because GOMS models are procedural and "under-constrain" behavior, there are many ways to implement a high-level command within the granular production system of ACT-R.
Mapping Representations
G2A maps GOMS operators—such as moving a mouse or clicking a button—to equivalent ACT-R production rules. While a GOMS model describes a sequence of steps, ACT-R functions as a collection of "if-then" rules triggered by internal buffers. G2A fills the gaps between these two paradigms by determining how high-level goals should be decomposed into low-level cognitive, visual, and motor operations.
Optimization via Hill-Climbing
To resolve ambiguities in how a task might be performed (e.g., different visual search strategies), G2A employs a hill-climbing search algorithm. The system generates candidate ACT-R models and evaluates their performance against a "target" duration, derived either from the original GOMS model or from pilot user data. By minimizing the mean squared error over hundreds of iterations, G2A identifies the most plausible model configuration for the task at hand.
Evaluation and Results
The researchers validated G2A using two primary benchmarks: Kieras’s Text Editing Model and empirical pilot user data.
Timing and Complexity
The system demonstrated remarkable accuracy in replicating the timing of the source models. G2A-generated ACT-R models matched the original GOMS durations with less than 5% error overall, and less than 1% error for total task duration. Furthermore, the system successfully managed the complexity of scale, expanding a 190-line GOMS model into a massive ACT-R production system containing 79 rules and over 1,500 lines of code.
Human Performance Prediction
When tuned using data from a single pilot user, the resulting model was used to predict the performance of six other users. The system achieved an error rate of approximately 35%. While this figure may seem high, it is consistent with the "average" performance of automatically generated cognitive models in the field, highlighting the inherent difficulty in capturing individual behavioral nuances.
Insights and Future Directions
The G2A project reveals that AI can effectively bridge the gap between abstract task descriptions and high-fidelity cognitive simulations. By automating the "Aha" moments of visual processing—such as determining how attention shifts during a search—the system moves cognitive modeling from a manual craft toward an automated engineering discipline.
Future work aims to refine the internal representations used by G2A. The authors suggest transitioning to PDDL (Planning Domain Definition Language) to enable more formal analysis of the cognitive search space. Additionally, researchers are exploring whether general heuristics can be identified to further optimize the search process, potentially reducing the reliance on extensive iterative simulations.
Summary updated