Sufficient Social Reasoning to Learn Productive Cycles of Cooperation

Social systems are often characterized by the diverse capabilities and needs of the individuals within them.In many settings, complementary capacities and needs present an opportunity for cooperation that can in-crease both individual outcomes and overall social welfare;  however, it may not always be the case thatdyadic cooperation leads to beneficial outcomes, as any two individuals may not have the specific capa-bilities and needs to complement each other.  That is,  the system lacks a coincidence of desires.  Rather,there may exist cooperative cycles in which cooperation can exist along a chain of individuals such thatevery individual provides and receives productive support.  We characterize such situations as complexsocial dilemmas and investigate,  using a learning-theoretic approach,  the preferences and reasoning ca-pabilities  that  are  sufficient  to  enable  the  emergence  of  productive  cooperation  cycles.   Specifically,  wecarry out computational experiments in the context of a multi-agent resource exchange game—the MarsColony game—and demonstrate how distinct utility functions, which correspond to varying levels of other-regarding preferences and social reasoning, lead to emergent cooperative cycles. We draw on existing workon cooperation in social dilemmas and multi-agent learning to motivate the choice of each utility function,which corresponds to selfish and other-regarding preferences, direct and indirect reciprocity, and causal at-tribution. The results show that standard other-regarding preferences are insufficient in our setting. Causalattribution, in tandem with other-regarding preferences, is shown to be an effective mechanism for learningproductive cooperative cycles

Chase McDonald - Second Year Paper