Abstract (eng)
Combinatorial auctions allow to simultaneously sell multiple heterogeneous items with interdependencies. Examples for such auctions are spectrum auctions, where the government sells licenses for bands of the radio spectrum, and procurement auctions, where the auctioneer buys services or goods. Another prominent example are auctions for sponsored search. Here a search engine company shows advertisements next to the search results on its website and the advertisers bid money for their ads being shown next to the search results for certain keywords.
One of the standard mechanisms for combinatorial auctions is the seminal VCG mechanism. This mechanism has the property that every bidder is incentivized to bid his valuation; this property is called incentive compatibility. Moreover, it has other desirable properties if bidders indeed bid their valuation: (1) No bidder has a negative utility; this property is called individual rationality. (2) The allocation of the items selected by the mechanism maximizes social welfare; that is, the allocation maximizes the sum of the valuations. However, these properties can only be achieved under certain conditions. It is necessary that (1) the utility functions of the bidders are quasi-linear, this implies that they cannot have budget constraints, and that (2) the computation of the social welfare maximizing allocation is possible in the available time. Thus, in settings where this is not the case it is not recommended to use the VCG mechanism. These are exactly the settings for which we design new auction mechanisms in this dissertation.
(A) First, we study settings where bidders are budget constrained. We design incentive compatible, individually rational, and Pareto-optimal mechanisms for auctions with heterogeneous items and bidders with additive valuations and budget constraints. (1) Specifically, we first study settings with single-dimensional valuations and prove a positive result for randomized mechanisms and an impossibility result for deterministic mechanisms. While the positive result allows for private budgets, the negative result is for public budgets.(2) Next, we study settings with multi-dimensional valuations. Here, we prove an impossibility result that applies to deterministic and randomized mechanisms even if budgets are public. Together with the previous results this shows the power of randomization in certain settings with heterogeneous items, but it also shows its limitations. (3) Furthermore, we study multiple keyword sponsored search auctions with budgets. Here, each keyword has multiple ad slots with so-called click-through rates. The bidders have single-dimensional additive valuations, which are linear in the click-through rates. Additionally, we assume that the number of slots per keyword assigned to a bidder is bounded. We give the first mechanism for multiple keywords and click-through rates differing among slots that is incentive compatible in expectation, individually rational in expectation, and that satisfies Pareto-optimality ex-post.
(B) Next, we study settings with complex valuations for which maximizing social welfare is NP-hard and is thus impracticable for large instances. We consider various classes of valuation functions and bidding functions and study the performance of the VCG mechanism when bidders are forced to choose their bids from a subclass of the class of valuation functions. Thus, the auctioneer restricts the class of bidding functions such that the VCG outcome can be computed efficiently. The performance of the mechanisms is measured in terms of the Price of Anarchy, which is the ratio of the maximum social welfare of all allocations and the minimum social welfare of outcomes when the bids form a strategic equilibrium. We show improved upper bounds on the Price of Anarchy for restrictions to additive bids and upper and lower bounds for restrictions to non-additive bids. Our bounds show that increased expressiveness can give rise to additional equilibria of poorer efficiency.
(C) Finally, we generalize the standard model for combinatorial auctions where the utility of a bidder depends solely on the item set assigned to him and on his payment. For instance, in online advertisement an advertiser might not want his ads to be displayed on the same page as the ads of his direct competitor. We propose and analyze several natural, simple graph-theoretic models for combinatorial auctions that incorporate negative, conflict-based externalities. Here bidders are embedded into a directed conflict graph, and D is the maximum out-degree of any node. We design O(D log log D / log D)-approximate algorithms and O(D)-approximate incentive compatible mechanisms for social welfare maximization. These ratios are almost optimal given existing hardness results for the independent set problem. For the prominent application of sponsored search, we present several algorithms when the number of items is small---the most relevant scenario in practice. In particular, we show how to obtain an approximation ratio that is sublinear in D when the number of items is only logarithmic in the number of bidders. All our algorithms for sponsored search can be turned into incentive compatible mechanisms.