Preference Elicitation for Risky and Intertemporal Choices

Speaker: Greg Hines, University of Waterloo

In many areas of artificial intelligence, we may be interested in making choices on behalf of users. We may also wish to provide the user with a worst case guarantee, such as minimax regret, of the performance of our choices. In order to do so, we must first query the user about their preferences. In this talk, I will examine cases when the choices involve an element of risk or time. I will show that current querying methods are not compatible with realistic models of preferences over risky choices, such as cumulative prospect theory (CPT). I present a new querying method that is compatible with both CPT and minimax regret. I also present heuristics for helping to determine the optimal query to ask the user. Finally, I show how to generalize this work to a temporal case where we must also query the user about their discount rates.