In group decision-making problems that involve self-interested agents with private information, reaching socially optimal outcomes requires aligning the goals of individuals with the welfare of the entire group. The well-known VCG mechanism achieves this by requiring specific payments from agents to a central coordinator. However, when the goal of coordination is to allow the group to jointly realize the greatest possible welfare, these payments amount to an unwanted cost of implementation, or waste. While it has often been stated that the payments VCG prescribes are necessary in order to implement the socially optimal outcome in dominant strategies without running a deficit, this is in fact not generally true. Cavallo (2006) specified the mechanism that requires the minimal payments among all mechanisms that are socially optimal, never run a deficit, and are ex post individual rational with an anonymity property. The mechanism achieves significant savings over VCG in a broad range of practically relevant domains, including allocation problems, by using information about the structure of valuations in the domain. This paper gives a high-level overview of that result, and discusses some potential applications to AI.
Subjects: 7.1 Multi-Agent Systems; 7. Distributed AI