Online Optimization for Markets and the Cloud

Abstract

Internet applications provide interesting dynamic environments for online optimization techniques. In this talk, I will discuss a number of such problems in the context of online markets, and in serving cloud services. For online markets, I discuss problems in online advertising. Online ads are delivered in a real-time fashion under uncertainty in an environment with strategic agents. Making such real-time (or online) decisions without knowing the future is challenging for repeated auctions. In this context, I will first highlight the practical importance of considering "hybrid" models that can take advantage of forecasting, and at the same time, are robust against adversarial changes in the input. In particular, I discuss our recent results combining stochastic and adversarial input models. Then I will present more recent results concerning online bundling schemes that can be applied to repeated auction environments. In this part, I discuss ideas from our recent papers about online bundling, stateful pricing, bank account mechanisms, and Martingale auctions. For problems on the cloud, I will touch upon two online load balancing problems: one in the context of consistent hashing with bounded loads for dynamic environments, and one in the context of multi-dimensional load balancing. Other than presenting theoretical results on these topics, we show how some of our new algorithmic techniques have been applied by Google and other companies, and confirm their significance in practice.