Abstract

Combinatorial Auctions (CAs) are promising to increase so-
cial welfare by enabling bidders to express their valuation
on any combination of items. A major issue of many CAs
is the requirement to optimally solve the NP-hard Combi-
natorial Allocation Problem. To release a centralized auc-
tioneer from that computational burden he can shift it to
the bidders. One of the few discussed decentralized auc-
tions is PAUSE, in which bidders suggest new allocations to
the auctioneer. In our theoretical analysis we examine the
bidders' bid complexity and determine a worst case bound
concerning eciency, if bidders follow a pro t maximizing
strategy. Based on these results we conduct computational
experiments with di erent bidding and computation strate-
gies, and analyze their impact on eciency, auctioneer's rev-
enue and auction runtime. Surprisingly, even if agents de-
viate from the optimal bid price calculation, PAUSE still
achieves high levels of eciency and auctioneer's revenue
compared to the Combinatorial Clock auction.

Share

COinS