Simulation to predict performance of assistive interfaces
Pradipta Biswas, Peter Robinson · 2007 · Proceedings of the 9th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS '07) · doi:10.1145/1296843.1296885
Summary
This short ASSETS 2007 poster from Pradipta Biswas and Peter Robinson at the Cambridge Computer Laboratory describes a simulator for predicting how disabled users will perform on a given assistive interface, intended to reduce the dependence on hard-to-recruit disabled participants for early-stage usability evaluation. The simulator takes a task definition and the locations of interface objects as input and outputs a predicted cursor trace and task-completion time for a chosen input device (mouse, single-switch scanning, etc.) and a chosen user profile (varying physical disability and skill level). It is structured as three coordinated models: an Application Model that decomposes the task into atomic sub-tasks, an Interface Model that selects the appropriate input modality and parameter set for the simulated user, and a User Model that drives the simulation through perception, cognition, and motor sub-models structured along the lines of Card, Moran and Newell's Model Human Processor (MHP). The cognitive sub-model uses CPM-GOMS to represent expert (optimal) behaviour and a probabilistic rule-based system for sub-optimal behaviour. The authors situate the work against earlier projects — AVANTI for adaptive web interfaces, McMillan's system-side modelling, and Keates et al.'s MHP-based comparison of able-bodied and motor-impaired users — and argue their approach is more general because it explicitly models perceptual, cognitive, and motor behaviour rather than just system adaptation.
Key findings
A case study used the simulator to predict task-completion times for an eight-directional polar scanning system used by people who control a computer through one or two switches. The cognitive model was validated by comparing predictions against actual times collected from eight able-bodied participants performing the same scanning task. Across most participants the simulator predicted completion time with an overall standard error below 3% and no statistically significant difference from observed times (paired two-tailed t-test, t = 0.31), with two outlier participants. The simulator was also used to compare three scanning techniques — eight-directional scanning, block scanning (iterative equal-area subdivision of the screen), and a new cluster-based scanning approach the authors developed — and found cluster scanning outperformed the other two on the simulated workload. The motor-behaviour sub-model had not yet been validated against motor-impaired users at the time of writing; the validation in the paper covered only the cognitive sub-model with able-bodied participants (under the assumption that the intended user population had no cognitive impairment).
Relevance
For accessibility researchers and tool builders, this paper is an early example of simulator-based evaluation of assistive interfaces — a methodology that has since become more important as machine learning enables richer user models, and as the field continues to grapple with the difficulty of recruiting representative disabled participants for early-stage interaction-technique comparisons. The three-model decomposition (application, interface, user) and the explicit borrowing from MHP and CPM-GOMS make the architecture relatively portable to other input modalities and tasks. The practical takeaway for designers of switch, eye-gaze, or scanning interfaces is that even a coarse simulator can rank competing techniques quickly enough to filter out clearly inferior designs before user testing. Limitations are substantial and the authors are upfront about them: the motor sub-model is unvalidated against the target population, the cognitive validation uses able-bodied users as a proxy, only one task class (eight-directional scanning) is tested, and the simulator depends on hand-built atomic-task decompositions rather than learning from logs. Readers should treat this as a pointer to the broader Biswas-Robinson research programme on simulator-based assistive-interface evaluation rather than as a fully validated tool.
Tags: user simulation · usability evaluation · GOMS · Model Human Processor · scanning interface · switch access · motor accessibility · human-computer interaction · cognitive modelling · assistive technology · accessibility evaluation