Welfare Engineering in Multiagent Systems

A multiagent system may be regarded as an artificial society of autonomous software agents. Welfare economics provides formal models of how the distribution of resources amongst the members of a society affects the well-being of that society as a whole. In multiagent systems research, the concept of social welfare is usually given a utilitarian interpretation, i.e. whatever increases the average welfare of the agents inhabiting a society is taken to be beneficial for society as well. While this is indeed appropriate for a wide range of applications, we believe that it is worthwhile to also consider some of the other social welfare orderings that have been studied in the social sciences. In this paper, we put forward an engineering approach to welfare economics in multiagent systems by investigating the following question: Given a particular social welfare ordering appropriate for some application domain, how can we design practical criteria that will allow agents to decide locally whether or not a proposed deal would further social welfare with respect to that ordering? In particular, we review previous results on negotiating Pareto optimal allocations of resources as well as allocations that maximise egalitarian social welfare under this general perspective. We also provide new results on negotiating Lorenz optimal allocations, which may be regarded as a compromise between the utilitarian and the egalitarian approaches. Finally, we briefly discuss elitist agent societies, where social welfare is tied to the welfare of the most successful agent, as well as the notion of envy-freeness.