Exploring Design Methods for Dynamic User Preferences

Abstract

New product development requires engineering designers to translate abstract user needs into concrete engineering specifications. Formal methods for engineering design offer strategies to guide decision-making throughout this process. The objective of these methods is to define product attributes which will maximize user preference relative to other options and therefore improve the likelihood of purchase. However, the emergence of the Internet of Things, smart products, and highly-connected systems has led to a growth in products sold as a service. Thus, user preferences during the entire product lifetime becomes increasingly important. Comprised of three studies, this dissertation will extend design research to investigate approaches which maximize user preferences over the lifetime of the product. The first study focuses on the design of large connected user-product-environment systems. Usage context and more specifically, the dynamic usage variables that are impacted indirectly by the designer’s decisions through the large-scale change in the end-users’ behavior. In the current literature, context variables are treated as uncontrollable variables. Designers’ approach to uncontrollable variables is developing predictive models to predict these variables accurately during the product life cycle and design the product accordingly. In this work, we propose a framework for bringing some dynamic aspects of the usage context into the controllable variable space which paves the way for designing the product and the usage context simultaneously. This will not only help with mitigating the negative impact of a product on its usage context, but will also offer a tool to change the usage context in order to get the most out of the product. Second study focuses on endogenous factors that change the user’s preference. We developed a framework using Reinforcement Learning to design products that are adaptive and able to change their attributes over time as they interact with the user. This helps with the design of a product that not only matches each user’s preferences (mass personalization), but also changes its attributes to maximize user satisfaction during the entire product life cycle without the need for designer’s input or any information about the cause of change. Using real data on preference change for the design of a variable stiffness prosthetic ankle, we explored different design approaches including traditional methods along with the RL framework, and compared different KPIs for the design approaches. Results show the superiority of the proposed framework over traditional design methods. In the last study, we follow the framework proposed in the second study and explore how different aspects of a reinforcement algorithm exploration/exploitation behavior impact user willingness for providing feedback to an adaptive engineered system. A pilot study of 29 participants was conducted using an adaptive office chair. Statistical analysis of the results shows that the desirability of the system impacts the reported user willingness to interact over long periods of time. However, experiment data did not support the hypothesis that responsiveness of the system makes a significant difference in user willingness compared to desirable unresponsive system states. Together these studies propose new frameworks for designing products that react to user preference changes as well as optimizing the system level impact of the product leading to better user satisfaction throughout the product life-cycle. The dissertation is concluded with a discussion on how these findings add to the current state of work along with the limitations and the implications of the results.