Miscomputing Ratio: The Social Cost of Selfish Computing

Auctions are useful mechanism for allocating items (goods, tasks, resources, etc.) in multiagent systems. The bulk of auction theory assumes that the bidders’ valuations for items are given a priori. In many applications, however, the bidders need to expend significant effort to determine their valuations. In this paper we analyze computational bidder agents that can refine their valuations (own and others’) using computation. We introduce a way of measuring the negative impact of agents choosing computing strategies selfishly. Our miscomputing ratio isolates the effect of selfish computing from that of selfish bidding. We show that under both limited computing and costly computing, the outcome can be arbitrarily far worse than in the case where computations are coordinated. However, under reasonable assumptions on how limited computing changes valuations, bounds can be obtained. Finally, we show that by carefully designing computing cost functions, it is possible to provide appropriate incentives for bidders to choose computing policies that result in the optimal social welfare.