My research project focuses on Conditional Preference Networks (CP-nets). These are graphical models for representing and reasoning with a person’s preferences over a set of discrete variables. The natural applications of these networks are in Artificial Intelligence as preference modelling is vital for applications such as recommender systems, automated decision making, and product configuration.
One of my first year contributions to this area was to create a new quantification for user preferences, given a CP-net representation. That is, I have used the semantics of CP-nets to construct a rank value for outcomes of interest. This value is an accurate representation of the user's level of preference for this outcome. Being able to quantify preference in this manner makes reasoning with preference information easier. In particular, these rank values can improve the efficiency of answering dominance queries. A dominance query asks, given a pair of outcomes, which will the user prefer? On the surface, this appears to be a simple question and one we would naturally like to be able to answer. However, dominance testing turns out to be a complex problem. Having a quantitative measure of preference allows us to enforce a numerical bound on the query, enabling us to answer it more efficiently. These results are in the process of being written up for publication.
As CP-nets are relatively new, there are many areas my project could explore. As my thesis title suggests, I intend to move on to looking at extensions of the basic CP-net. Many extensions have been proposed, however I shall be focusing particularly on PCP-nets, mCP-nets, and UCP-nets. PCP-nets are CP-nets which also incorporate probability in the framework. They allow us to express uncertainty over the user’s preferences. mCP-nets are structures which incorporate the preferences (CP-nets) of multiple agents into one framework. Finally, UCP-nets specify not just preferences but conditional utilities over the variables of interest. This provides more information about the user’s preferences but should not be too much harder to elicit from a non-expert.
- MMath, BSc Mathematics - University of Leeds
Research groups and institutes