Theoretical and Experimental Insights into Decentralized Combinatorial Auctions

Combinatorial Auctions (CAs) are promising to increase social 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 Combinatorial Allocation Problem. To release a centralized auctioneer from that computational burden he can shift it to the bidders. One of the few discussed decentralized auctions 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 efficiency, if bidders follow a profit maximizing strategy. Based on these results we conduct computational experiments with different bidding and computation strategies, and analyze their impact on efficiency, auctioneer’s revenue and auction runtime. Surprisingly, even if agents deviate from the optimal bid price calculation, PAUSE still achieves high levels of efficiency and auctioneer’s revenue compared to the Combinatorial Clock auction.