EAGER: Preferences in Repeated Choices

University Of Kentucky Research Foundation, Lexington KY

Investigators

Abstract

Computational models of preferences are important for recommendation and decision-support systems, with applications including long-term planners, smart homes, and health-care monitors. Because there is a trade-off between the expressivity of preference representations and the ease of reasoning with them, most personalized preference-based systems rely on either statistical models of similar individuals, or very simple models of one individual?s preferences. We propose to work with conditional preference models, which can represent more complex preferences. Our work will be to develop better algorithms for reasoning about preferences over complex outcomes or scenarios, given such models, and also to extend the model to be able to reason about repeated choices over time. In addition, the PI will use preference modeling software in her outreach to schools, colleges, and the general public to show ways that computers can model and reason about preferences. The PI has a record of supporting a diverse group of students, including students from the LGBTQA community, students with learning disabilities, women, and people of color. Thus, the broader impacts include outreach and a software package, which will be publicly available, for outreach about AI and preference handling, contributions to the infrastructure of preference reasoning research, and support of diversity in computer science. The first part of our proposal is to work algorithms for deciding, given a conditional preference network (CP-net) and two outcomes, which outcome is more preferred. We will organize a competition, perhaps based on the ICAPS International Planning Competitions, for these algorithms. Secondly, we will explore two models of temporal preferences: Temporal Conditional Choice Networks (TCC-nets) and hidden Markov models (HMMs). We will develop an iPhone app, CommuteRoute, to collect individuals? choices over time of routes between home and work. This data, stored as feature vectors, will (in future work) allow us to test algorithms for learning and reasoning with TCC-nets, and comparing those algorithms to extant HMM algorithms.