A First Approach to Autonomous Bidding in Ad Auctions

We present autonomous bidding strategies for ad auctions, first for a stylized problem, and then for the more realistic Trading Agent Competition for Ad Auctions (TAC AA)—a simulated market environment that tries to capture some of the complex dynamics of bidding in ad auctions. We decompose the agent’s problem into a modeling subproblem, where we estimate values such as click probability and cost per click, and an optimization sub-problem, where we determine what to bid given these estimates. Our optimization algorithms, the focus of this paper, come in two flavors: rule-based algorithms, which can make reasonable decisions even with inaccurate models; and greedy multiple choice knapsack algorithms, which can make better decisions but require more accurate models to do so. We evaluate agent performance in the game-theoretic TAC AA domain, as well as in a more controlled testing framework that decouples the modeling and optimization subproblems. Although decision-theoretic in nature, we argue that it is nonetheless reasonable to use our controlled testing framework to determine which optimization algorithm is most suitable for an agent with a given amount of model error, and which model improvements would lead to the greatest gain in overall agent performance.