Empirical Bi-Action Tables: A Tool for the Evaluation and Optimization of Text-Input Systems. Application I: Stylus Keyboards

We introduce a technique that, given any text input system A and novice user u, will predict the peak expert input speed of u on A, avoiding the costly process of actually training u to expert level. Here, peak refers to periods of ideal performance, free from hesitation or concentration lapse, and expert refers to asymptotic competence (e.g., touch typing, in the case of a two-handed keyboard). The technique is intended as a feedback mechanism in the interface development cycle between abstract mathematical modeling at the start (Fitts' law, Hick's law, etc.) and full empirical testing at the end. The utility of the technique in iterative design is contingent on what we call the monotonicity principle: For each user u, if our prediction of peak expert input speed for u is higher on system A than on system B, continuous text input by u after training to expert level will be faster on A than on B. Here, continuous refers to actual real-world use, subject to errors, physical fatigue, lapses of concentration, and so forth. We discuss the circumstances under which monotonicity is valid. The technique is parametric in the character map-that is, in the map from actions (keystrokes, gestures, chords, etc.) to characters. Therefore, standard heuristic algorithms can be employed to search for optimal character maps (e.g., keyboard layouts). We illustrate the use of our technique for evaluation and optimization in the context of stylus keyboards, first benchmarking a number of stylus keyboards relative to a simple alphabetic layout and then implementing an ant algorithm to obtain a machine-optimized layout.