RI: Small: Agent-Assisted Trading in Real-World Auctions

Brown University, Providence RI

Investigators

Abstract

In many large economic markets, goods are sold through auctions. Examples include eBay, wireless spectrum auctions, the New York Stock Exchange, and the Dutch flower auctions. This project focuses on one particular real-world auction domain: the Dutch flower auctions (DFA). The project will perform an in-depth analysis of this domain both from the point of view of the bidders and from the point of view of the auctioneers. This project involves several component tasks. (1) Creation of a configurable auction server, comprised of basic auction building blocks, like simultaneous and sequential auctions, and sealed-bid, ascending, and descending auctions. (2) Construction of an interactive DFA simulation using the configurable auction server that is capable of handling decisions made by both human bidders and autonomous agents; this simulation will model DFA bidder preferences, which will be learned automatically using inverse reinforcement learning techniques from historical data and manually by surveying and interviewing DFA experts. (3) Creation of autonomous DFA bidding agents and comparison of their performance to that of expert human DFA bidders using the interactive simulation. (4) Further simulation of these agents to empirically find game-theoretic equilibria under various settings of the auction parameters so that those parameters can be optimized. The project will also develop a prototype of a mixed-initiative system that can be used to assist DFA bidders. Potential broader impacts of this research include increased knowledge and understanding of how to design and implement artificially intelligent agents that can effectively assist humans with their decision-making efforts, particularly in information-rich and time-critical environments. Although this project focuses on DFA-type auctions, methodologies developed here will transfer to other domains that use auction mechanisms, such as automated bidding for resource allocation in smart grids. Detailed focus on the DFA provides a concrete starting point for the study of various auction-related topics, for instance, how to build mixed-initiative decision support tools to improve the quality of bidding practices, rigorous comparison of various auction mechanisms and parameter choices, and advancing bidding agent design from a heuristic approach to one that is theoretically grounded. The project research team includes both graduate and undergraduate students whose participation in this research will strengthen their understanding of the various fields that are critical to the development of decision support systems (e.g., autonomous agents, preference elicitation and representation, machine learning, software engineering) and provide them experience collaborating across disciplines, and working on a real-world application. In addition, there are plans to develop introductory lessons on the general topic of autonomous bidding agents for use in the Artemis project, a five week summer program in which female rising ninth graders are exposed to the breadth of applications of computer science, and are introduced to a variety of technologies underlying computing.