 Hello Internet, this is Helming from Rensselaer Polytechnic Institute. In this video, I'm gonna talk about our research on learning people's preferences, in particular by asking them to compare candidates. Well, this doesn't sound like too hard of a problem, right? I mean, I can just ask you two questions, maybe three. Or how about I just ask you to rank them all together? One is, what if people's preferences are non-deterministic? That is, your preference today may be different from that of yesterday. Makes sense, right? Because even if you're sure that you like McDonald's more than Burger King, it doesn't mean you will go to McDonald's every time you have a choice between the two. It's just you will go to McDonald's more often. Also what if, instead of choosing a place to eat by yourself, you are eating out with a group of friends and you want to make a group decision? What if, instead of just three burger places, I want to learn your preferences over all the restaurants in your area, some of them you haven't been to yet? In the end, I want to be able to recommend you some new places, without bothering you with too many questions. Here's where our elicitation algorithm comes into play. We ask for pairwise comparisons that maximizes information gain. From the set of restaurants you've been to, we're gonna pick two for you to compare so that we can learn a batch of overall preferences as much as possible. In the algorithm, we'll utilize a relatively newly proposed information criterion, which measures the information gain from your response by the minimum certainty of the comparison between any two candidates. We did some experiments with synthetic data. From the figure above, we can see that as we ask more and more questions using this information criterion rather than others, it's more likely to recommend candidate that actually gives you high utility. This algorithm or its methodology can be used to ask a larger variety of elicitation questions, such as to rank or to pick the best from a set of 10 candidates. From there, we can go a step further and ask questions that maximizes information gain subtracted by the cause of answering such questions. For example, asking you to rank 10 restaurants can tell us a lot about your preferences, but it might take you quite some effort to rank them all. Instead, it might be better to just ask for your favorite one from the 10. That's all. I hope you enjoyed this brief review of our research at the Cognitive and Immersive Systems Lab here at RPI. I want to thank you on behalf of my teammates, Jirving and Professor Lirong. Hope to see you soon.