Social and Asocial Learning in Collective Action Problems: The Rise and Fall of Socially-Beneficial Behaviour

The allocation of common-pool resources is an important topic in technical and socio-technical systems, and when left unmanaged, such systems often collapse to highly unequal and unsustainable outcomes. Recent work has highlighted a role for electronic institutions in managing such resources, to ensure socially-beneficial outcomes in the long term. However, open self-organising multi-agent systems often involve agents that learn behaviours in order to meet their goals. In this paper we explore the interplay between institutional features and forms of social and asocial learning employed by self-interested agents. We show that, while recent results have associated social learning with sustainability, this is sensitive to the form of social learning used. We show that more realistic models that combine social and asocial learning are more likely to lead to unsustainable institutions and anti-social outcomes. However, a key role for pardons in the sanction mechanism of the institution is identified, which allows for tolerance of a range of behaviours associated with ongoing learning, including complacency and exploration.