Computer Science Events
[PAST EVENT] Decision Making FOR Groups and BY Groups
Access & Features
- Open to the public
Speaker: Dr. Sujoy Kumar Sikdar from Washington University in St. Louis
Fundamental societal problems often require making decisions for groups of agents over multiple attributes simultaneously, such as allocating or dividing public and private resources of multiple types, or making decisions on multiple issues simultaneously, when the agents have complex preferences over the alternatives. Examples of such problems are ubiquitous: assisting the homeless and vulnerable members of society through the allocation of public resources, allocating papers and time slots in a seminar class, allocating computing resources in cloud computing, selecting members of a committee, shortlisting candidates for a job, and voting over multiple ballot measures. At the same time, recent advances in communications have led to an abundance of information about human preferences and the efficacy of previous decisions, either explicitly, or more frequently, implicitly through our actions and communications. My research addresses both of these aspects of group decision making: Making fair and economically efficient decisions given agents’ preferences; and representing and learning preferences from real world data on group decisions.
In this talk, I place a spotlight on two recent pieces of work that address these research agendas. First, I focus on the problem of multi-type housing markets, where I provide mechanisms for facilitating the distribution of multiple types of items that are initially owned by individuals in a group, who have preferences over bundles that consist of combinations of the different types of items. I describe a mechanism that (1) always produces a core allocation: a redistribution of items with the guarantee that no group of agents can benefit by redistributing only the items they own, only among themselves, and (2) is immune to any group of strategic agents manipulating their preferences in order to obtain a better allocation. This result breaks a long-standing impossibility result in the design of core allocation selecting and strategyproof mechanisms for multiple types of items, by imposing a natural restriction on the domain of agents’ preferences. In the process I discuss a novel preference representation language that generalizes several previously studied representations. In the second half of the talk, I describe an exciting new approach to improve the usability of online voting and preference elicitation platforms by learning about agents’ preferences from past activity, and recommending an initial ordering that minimizes the expected time a user will have to spend in submitting their true preference. Finally, I outline how we may in the near future, develop methods to elicit preferences, make fair and efficient decisions, measure outcomes, and improve decision making.
Bio: Sujoy Sikdar is a postdoc at the Department of Computer Science and Engineering at the Washington University in St. Louis (WashU). He received his Ph.D. and MS in Computer Science at the Rensselaer Polytechnic Institute (RPI). He has broad interests across artificial intelligence, mechanism design, machine learning, and computational social science. His research focuses on algorithmic decision making in the allocation of resources, and voting, with an eye at learning and faithfully representing preferences. He is the recipient of a best paper award at SocialCom2013, and his dissertation was nominated by RPI for the Joint AAAI/ACM SIGAI Doctoral Dissertation award.
Contact
Zhenming Liu