Dynamics of Strategic Manipulation in AdWords Auction
Tian-Ming Bu
2007-01-01
Abstract:The rise of Internet advertisement has created a demand for new auction models. The ad-words auction, used by Google for this purpose, called the position auction by Varian [15], assigns advertisement slots to top price bidders in a decreasing order of the attractiveness to the viewers, and charge each winning agent, for every click generated from its weblink at the allocated slot, the price of the next lower bid. It has been known that this protocol is not incentive compatible but a pure Nash equilibrium exists for the game among the bidders. Moreover, refined solution concepts based on Nash equilibrium have been proposed for better understanding this game [9, 15]. As the participating agents may not bid their private values in the game, it is not clear whether those existential equilibria can be reached. In this paper, we are interested in the dynamic process how bidders interact to reach an equilibrium. We first propose a new solution concept, the forwardlooking Nash Equilibrium, for the position auction by considering the strategic manipulations of an agent that take into consideration the effect of the existing strategies of other agents, as well as their future responses to its own benefit. We prove that the forward-looking Nash equilibrium has a unique solution. We note that, in a Nash equilibrium, bidders with higher private values may be stuck with an allocation worse than others with a lower private value of the allocation. The forward-looking Nash equilibrium, on the other hand, guarantees a concept called output truthful in that the higher a bidder’s true value is, the better allocation it will obtain. Interestingly, we prove that the forward-looking Nash equilibrium in its pricing and allocation scheme is equivalent to the VCG auction outcome, which was regarded not suitable as too complicated to transfer into it from the current Google’s position auction pricing system. In fact, our results justify the use of Google’s position auction pricing scheme. In presence of the most sophisticated users, it is indeed equivalent to the Vickrey auction pricing scheme. Most importantly, we justify the new solution concept by deriving its convergence property. We study several dynamic adjustment schemes by the bidders, including one ∗The work was done while the author was at City University of Hong Kong. Research is supported by a research grant of National Natural Science Foundation of China (No. 60496321) Copyright is held by the author/owner(s). WWW2007, May 8–12, 2007, Banff, Canada. . that converges in a finite number of steps. Moreover, we show that a randomized adjustment scheme will converge to the forward-looking Nash equilibrium with probability one. As the randomized scheme is most reasonable to characterize a situation of uncoordinated bidders, it shows robustness of the forward-looking Nash equilibrium.