Supporting Sustainable Decision-Making with Value-Sensitive Artificial Intelligence

Abstract

Sustainability is a term that is becoming increasingly prevalent, as several recent catastrophic events are often attributed to the impact that modern lifestyle has on the environment. Moreover, the term extends also to sustainability of operational democratic societies, where availability and accessibility to several critical infrastructures is ensured for individuals. Yet, achieving sustainability can be challenging. Recent reports from the United Nations conclude that improving decision-making processes on several levels, from institutional level policy making to individual level everyday decisions, supports sustainable development. Typically, deciding on more "sustainable" solutions often leads to complex multi-objective optimization problems, which are not trivial to solve. Modern artificial intelligence (AI) provides several methods to handle complex problems, particularly within the fields of machine learning and optimization. Nevertheless, AI has also given rise to new challenges, especially related to privacy, autonomy, and personal values and morals. Thus, the need for value-sensitive AI systems arises, where values and preferences are included directly in the system design. The current thesis pursues the paradigm of efficient value-sensitive AI, mainly focusing on sustainable decision-making problems, by providing experimental, empirical, and methodological arguments. Three main design approaches are considered: centralized, decentralized, and a hybrid combination of both. First, an AI system applies centralized controls to simulations of critical infrastructure components, such as power grids. The results show the ability of said AI controls to stabilize and sustain the operation of critical infrastructures. Yet, centralized approaches may restrict and suppress personal freedoms and autonomy, especially when applied on individuals. Thus, a decentralized approach is evaluated next in the domain of sustainable product recommendations. The proposed AI system receives explicit input from individuals regarding their morals and values, and then calculates personalized ratings for sustainable products. Interactions between the decentralized system and individuals are evaluated in a field-study on two grocery stores. Statistically significant results confirm the system's ability to support individuals towards more sustainable purchases. Finally, hybrid combinations of centralized and decentralized approaches are evaluated. A novel privacy-preserving framework is proposed to calculate accurate aggregations of individual data without exposing the actual individual data to centralized systems. Additionally, a novel hybrid AI system is introduced and combined with the privacy-preserving framework to generate sustainable basket recommendations based on personal values and environmental objectives. Quantitative results on a synthetic dataset show a considerable reduction of environmental impact, even when users adopt only a fraction of the recommendations.