A modular framework for multi-agent preference elicitation

I present a framework for multi-agent preference elicitation in the context of a discrete resource-allocation problem, known as the combinatorial allocation problem (CAP). There are several distinct, indivisible items, which must be allocated among a set of agents. The agents value bundles rather than just individual items. Because the number of bundles can be very large, agent preferences cannot be exhaustively described. An elicitation scheme for the CAP must therefore carefully choose the language in which it will model agent preferences to ensure succinct representations. The approach I propose is to embed learning algorithms for certain preference representations into the resource-allocation process. Preferences are elicited incrementally, and at well-defined breakpoints a tentative allocation is computed. This process is repeated to the extent needed until an efficient allocation is found. The framework is modular in that a variety of different learning algorithms can be introduced as subroutines to construct models of the individuals agents' preferences, as long as the subroutines interact with the agents through a standard query interface. The current leading distributed algorithms for the CAP are iterative combinatorial auctions, but the iterative combinatorial auctions that can guarantee allocatioe efficiency all use a single bidding language, namely XOR, that may not be appropriate for certain applications. Experimental results demonstrate that the elicitation framework can complement current designs by allowing for alternate representations where XOR is inappropriate, resulting in fewer queries and faster convergence. The framework consists of elicitation, allocation, and pricing engines. The pricing engine is also modular. I present two different methods for pricing, one of which can also serve as a stand-alone iterative auction. The auction begins with item prices and introduces bundle prices as needed to drive the bidding forward. This again complements existing designs which are limited to XOR or item pricing. The framework can also be extended to compute VCG payments, to bring truthful responses to queries into an equilibrium.