Nic Wilson: On the Problem of Optimising with Respect to Partially Known Multi-Objective User Preferences

Tuesday March 19 2024 at 15:00am in room B304

Speaker: Nic Wilson (University College Cork, Ireland)

Title: On the Problem of Optimising with Respect to Partially Known Multi-Objective User Preferences

Summary: Commonly, a decision support system will have only partial information about the preferences of a user. In multi-objective optimisation problems, user preferences are often modelled with some form of weighted sum of objective values, where the vector of weights is initially unknown.  In this talk I will consider reasoning and optimisation with partial knowledge of preferences for such multi-objective optimisation problems.  Throughout, my focus will be on fundamental concepts rather than algorithmic techniques. 

 If an alternative is necessarily optimal, i.e., optimal for each consistent candidate preference function, then the system can safely recommend that alternative (even though the knowledge of the user preferences may still only be partial). Otherwise, there are different potentially optimal alternatives. One can increase the knowledge of the user preferences by querying the user, asking the user to compare two or more alternatives. Max regret is a powerful tool which can be used in the generation of queries, as well as for generating a (pessimistic) recommendation, and in a stopping criterion; however, it is sensitive to how the objectives are expressed, which suggests a lack of robustness. I will discuss querying strategies, which aim towards finding a necessarily optimal alternative, including based on max regret.