Large-scale hierarchical optimization for online advertising and wind farm planning
thesisposted on 01.08.2013 by Konstantin Salomatin
In order to distinguish essays and pre-prints from academic theses, we have a separate category. These are often much longer text based documents than a paper.
This thesis develops a framework to investigate and design novel optimization methods for two important problems: computational advertising (particularly, sponsored search) and wind farm turbine-layout planning. Whereas very different in specifics, both problems share some common abstractions. The existing solution in sponsored search is based on a greedy pay-per-click auction and is suitable only for advertisers seeking a direct response. It does not apply to advertisers who target certain numbers of clicks in a predefined time period. To address this new challenge, we introduce a unified optimization framework combining pay-per-click auctions and guaranteed delivery in sponsored search. Our new method maximizes the revenue of the search engine, targets a guaranteed number of ad clicks per campaign for advertisers willing to pay a premium, and enables keyword auctions for all others. Results combining revenue to the search engine and click rates for the advertisers show superior performance over strong baselines. The proposed framework is based on linear programming with delayed column generation for computational tractability at scale. We design a game theoretic approach to optimize the strategy for individual advertisers, i.e. to optimize their choices between auctions and guaranteed delivery, and analyze the behavior of the new market formed by our framework. Specifically, we introduce a new method for computing the approximate Nash equilibrium where an exact computation would prove computationally intractable. We rely on approximations of complex utility functions, a combination of simulated annealing and integer linear programming as our principled approach. Wind farm layout optimization is the selection of optimal locations for placement of large wind turbines taking into account factors such as topographical features, prevalent but non-constant wind direction and turbine-wake interference. Existing approaches are deficient in their inability to consider long distance turbine interference, changing wind speed and direction and multiple types of wind turbines in optimization. The dissertation develops an optimization framework based on a scalable divided-and-conquer strategy that enables scalability to real-world wind farm scales taking into account the aforementioned complexities in the optimization process. Essentially the process optimizes in a hierarchical manner at different levels of granularity. This hierarchical decomposition approach to optimization is common to both search-advertisement and wind-farm layout challenges.