B1793
Title: Modelling cooperation in a dynamic common pool resource game with deep reinforcement learning
Authors: Simon Gero Haastert - University of Münster (Germany) [presenting]
Matthias Hettich - TU Berlin (Germany)
Abstract: The rapid degradation of natural common pool resources such as common fisheries, rain forests, or the global greenhouse gas budget is one of the greatest threats to human well-being. Such non-excludable, finite resources are exposed to socially inadequate appropriation strategies - a phenomenon often termed the tragedy of commons. However, game theory, as well as many lab experiments and real-life case studies, assert that cooperation towards the sustainable use of a common-pool resource is feasible. Certain conditions like reciprocal punishment or reward mechanisms, preferences such as inequity-aversion, or ways of communication can facilitate cooperation. We propose voting on an ad valorem tax rate as another instrument to enable multiple agents to coordinate their appropriation efforts. We model the common pool resource game with an agent-based model. Agents act individually in a partially observable Markov game in a trial-and-error fashion. They influence each other by reducing the common resource stock and voting on a tax rate. We train the agents via deep reinforcement learning with a state-of-the-art algorithm, which allows a continuous state and action space. In this setting, agents learn to indirectly punish agents for above-average use and over-exploitation of the common-pool resource by voting for a non-zero tax rate. We find that introducing the tax voting mechanism facilitates cooperative behavior and improves both resource sustainability and social welfare.