Strategy and mechanism lessons from the first ad auctions trading agent competition
Proceedings of the 11th ACM conference on Electronic commerce
Probabilistic analysis of simulation-based games
ACM Transactions on Modeling and Computer Simulation (TOMACS)
Strategic analysis with simulation-based games
Winter Simulation Conference
| Hi-index | 0.00 |
As agent technology matures, it becomes easier to envision electronic marketplaces teeming with autonomous agents. Since agents are explicitly programmed to optimally compete in these marketplaces (within bounds of computational tractability), and markets themselves are designed with specific objectives in mind, tools are necessary for systematic analyses of strategic interactions among autonomous agents. While traditional game-theoretic approaches to the analysis of multi-agent systems can provide much insight, they are often inadequate, as they rely heavily on analytic tractability of the problem at hand; however, even mildly realistic models of electronic marketplaces contain enough complexity to render a fully analytic approach hopeless. To address questions not amenable to traditional theoretical approaches, I develop methods that allow systematic computational analysis of game-theoretic models in which the players' payoff functions are represented using simulations (i.e., simulation-based games). I develop a globally convergent algorithm for Nash equilibrium approximation in infinite simulation-based games, which I instantiate in the context of infinite games of incomplete information. Additionally, I use statistical learning techniques to improve the quality of Nash equilibrium approximation based on data collected from a game simulator. I also derive probabilistic confidence bounds and present convergence results about solutions of finite games modeled using simulations. The former allow an analyst to make statistically-founded statements about results based on game-theoretic simulations, while the latter provide formal justification for approximating game-theoretic solutions using simulation experiments. To address the broader mechanism design problem, I introduce an iterative algorithm for search in the design space, which requires a game solver as a subroutine. As a result, I enable computational mechanism design using simulation-based models of games by availing the designer of a set of solution tools geared specifically towards games modeled using simulations. I apply the developed computational techniques to analyze strategic procurement and answer design questions in a supply-chain simulation, as well as to analyze dynamic bidding strategies in sponsored search auctions. Indeed, the techniques I develop have broad potential applicability beyond electronic marketplaces; they are geared towards any system that features competing strategic players who respond to incentives in a way that can be reasonably predicted via a game-theoretic analysis.