Proposal April, 2000
Full Dissertation May, 2001
Classic Mechanism Design
Computational Mechanism Design
Linear Programming and Auction Design
iBundle: An Iterative Combinatorial Auction
Linear Programming and Vickrey Payments
iBundle Extend & Adjust
Bounded-Rational Compatible Auctions & Myopic Best-Response
Extended Example: Distributed Train Scheduling
A fundamental problem in building open distributed systems is to design
mechanisms that compute optimal system-wide
solutions despite the self-interest of individual users and
computational agents. Classic game-theoretic solutions are often
prohibitively expensive computationally.
For example, the Generalized Vickrey Auction (GVA) is an
efficient and strategy-proof solution to the combinatorial allocation problem
(CAP), in which agents demand bundles of items, but
every agent must reveal its value for all possible bundles and the
auctioneer must solve a sequence of NP-hard optimization problems to
compute the outcome.
I propose iBundle, an
iterative combinatorial auction in which agents can bid for
combinations of items and adjust their
bids in response to bids from other agents. iBundle computes the
efficient allocation in the CAP when
agents follow myopic best-response bidding strategies,
bidding for the bundle(s) that maximize
their surplus taking the current prices as fixed.
iBundle solves problems without complete information revelation from
terminates in competitive equilibrium.
Moreover, an agent can follow a myopic best-response strategy with approximate
values on bundles, for example with lower- and upper- bounds.
My approach to iterative mechanism design decomposes the problem into
two parts. First, I use linear programming theory to
develop an efficient iterative auction under the assumption that
agents will follow a myopic best-response bidding strategy. Second, I
extend the approach
to also compute Vickrey payments at the end of the auction. This
makes myopic best-response a sequentially-rational strategy for agents
in equilibrium, inheriting many of the useful game-theoretic
properties of the GVA.
iBundle implements a primal-dual algorithm, CombAuction,
for the CAP, computing a feasible primal (the
provisional allocation) and a feasible dual (the ask prices) that
satisfy complementary slackness conditions.
An extended auction, iBundle
Extend & Adjust, interprets a primal-dual algorithm,
VickAuction, as an iterative auction. VickAuction computes
the efficient allocation and Vickrey payments with only best-response
information from agents.
Experimental results demonstrate that
iBundle Extend & Adjust, which
keeps iBundle open for a second phase before adjusting prices
towards Vickrey payments, computes Vickrey payments across a suite of