Friday, August 4, 2017 - 2:00pm
Groseclose 402
Title: | The Price of Deception in Social Choice |
Advisor: | Prof. Craig Tovey, School of Industrial and Systems Engineering |
Committee: | Prof. Paul Griffin, School of Industrial Engineering, Purdue University |
Prof. Santosh Vempala, School of Computer Science | |
Prof. Prasad Tetali, School of Mathematics | |
Prof. Vijay Vazirani, School of Computer Science | |
Reader: | Prof. Dr. Jörg Rothe, Institute for Computer Science, University of Düsseldorf |
SUMMARY: Most social choice algorithms rely on private data from individuals to maximize a social utility. However, many algorithms are manipulable – an individual can manipulate her reported data to obtain an outcome she prefers often at the cost of social good. Literature addresses this by declaring an algorithm as “manipulable” or “strategy-proof”. However, for many decision we are forced to either use a manipulable algorithm or an algorithm with negative properties; for instance, a dictatorship is the only reasonably strategy-proof way to decide an election. Thus, we view it as unwise to take an all-or-nothing approach to manipulation since we often have to settle for nothing. In this dissertation, we focus on algorithmic design with strategic behavior in mind. Specifically we develop the framework to examine the effect of manipulation on social choice using a game theoretic model. Our model of human behavior is supported by an extensive amount of experimental evidence and psychology and economics and allows us to better understand the likely outcomes of strategic behavior. We introduce a measure of manipulation – the Price of Deception – that quantifies the impact of manipulation. With the Price of Deception we are able to identify algorithms are negligibly impacted by manipulation, algorithms where strategic behavior leads to arbitrarily poor outcomes, and anything in-between. We prove the power of our model and the Price of Deception by answering open problems in assignments, facility location, election and stable marriage including a 28-year open problem by Gusfield and Irving. Our results demonstrate that the Price of Deception, like computational complexity, provides finer distinctions of manipulability than between “yes” and “no”.