Simulating the Human in Human-Centred Design: Increasing Efficiency in Design Activities

Abstract

"All simulations are wrong, but some are useful." This quote is adapted from Prof. George E. P. Box's famous aphorism and encapsulates the central hypothesis of this thesis: well-structured simulation approaches, while inherently imperfect, can systematically enhance the efficiency and rigour of human-centred design (HCD), particularly particularly when access to stakeholders is limited and interdisciplinary complexity is high, such as designing augmentative and alternative communication (AAC) systems. This research is motivated by critical limitations observed in the practical application of the Design Research Methodology (DRM), a widely adopted model for structuring design research across four sequential stages: Research Clarification (RC), Descriptive Study I (DS-I), Prescriptive Study (PS), and Descriptive Study II (DS-II). Despite DRM’s conceptual clarity, its implementation often breaks down in real-world HCD and AAC contexts. The RC stage is time-consuming; DS-I and PS struggle with empirical grounding when user access is limited; and DS-II is inconsistently executed, particularly in domains requiring long-term validation. These issues are compounded in the era of generative AI, where user needs and technological implications are often under-defined or unstable at early stages. To address these limitations, this thesis introduces the triple-diamond design process, a staged, simulation-driven extension of DRM. This process unfolds in three phases—computer-simulation-based, proxy-user-based, and end-user-centred design—and strategically aligns with the corresponding DRM stages. Through three case studies situated in the design of AAC systems, the thesis explores how computational agents, surrogate user models, and structured transitions between design phases can be used to improve the efficiency and traceability of HCD. The first case study uses LLM-based agents to simulate stakeholder reasoning, supporting rapid insight generation during RC and DS-I. The second builds an imperfect surrogate user to explore interaction strategies and performance trade-offs in predictive text systems. The third case study integrates proxy-user evaluation and end-user feedback to iteratively refine a multimodal AAC system, completing the full DRM cycle. Together, these contributions offer empirical support for the staged simulation-driven process as a practical augmentation to DRM. This work demonstrates how simulations can reduce time and resource demands, compensate for missing data, and maintain empirical coherence in the design of complex human–AI systems.