Optimal Real Time Bidding in Online Advertising

Sammanfattning: This thesis explores some of the possibilities of demand side optimization in online advertising, specifically how to evaluate and bid optimally in real time bidding. Theory for many types of optimizations is discussed. The thesis evaluates auctions from a game theory and control theory perspective. It also discusses how big data sets can be used in real time, and how agents can explore unknown stochastic environments. All items are valued through an estimated action probability, and a control system is designed to minimize the cost for these actions. The control system aims to find the lowest possible price per item while spending the entire budget. Periodic market changes and censored data makes this task hard and imposes low pass characteristics on the closed system. Using data to evaluate items is a high dimensional problem with very small probabilities. When data is limited the algorithm is forced to choose between low variance and precision. The choice between exploring and exploiting the unknown environment is crucial for long and short term results. An optimization algorithm was implemented and run in a live environment. The algorithm was able to control the spend optimally, but distributed it suboptimally.