AAAI Publications, Twenty-Fourth AAAI Conference on Artificial Intelligence

Font Size: 
Can Approximation Circumvent Gibbard-Satterthwaite?
Ariel D. Procaccia

Last modified: 2010-07-04

Abstract


The Gibbard-Satterthwaite Theorem asserts that any reasonable voting rule cannot be strategyproof. A large body of research in AI deals with circumventing this theorem via computational considerations; the goal is to design voting rules that are computationally hard, in the worst-case, to manipulate. However, recent work indicates that the prominent voting rules are usually easy to manipulate. In this paper, we suggest a new CS-oriented approach to circumventing Gibbard-Satterthwaite, using randomization and approximation. Specifically, we wish to design strategyproof randomized voting rules that are close, in a standard approximation sense, to prominent score-based (deterministic) voting rules. We give tight lower and upper bounds on the approximation ratio achievable via strategyproof randomized rules with respect to positional scoring rules, Copeland, and Maximin.

Full Text: PDF