Analysing public goods games using reinforcement learning: effect of increasing group size on cooperation.

Abstract

Electricity competition, restrictions on carbon dioxide (CO2) emissions and arm races between nations are examples of social dilemmas within human society. In the presence of social dilemmas, rational choice in game theory leads to the avoidance of cooperative behaviour owing to its cost. However, in experiments using public goods games that simulate social dilemmas, humans have often exhibited cooperative behaviour that deviates from individual rationality. Despite extensive research, the alignment between human cooperative behaviour and game theory predictions remains inconsistent. This study proposes an alternative approach to solve this problem. We used Q-learning, a form of artificial intelligence that mimics decision-making processes of humans who do not possess the rationality assumed in game theory. This study explores the potential for cooperation by varying the number of participants in public goods games using deep Q-learning. The simulations demonstrate that agents with Q-learning can acquire cooperative behaviour similar to that of humans. Moreover, we found that cooperation is more likely to occur as the group size increases. These results support and reinforce existing experiments involving humans. In addition, they have potential applications for creating cooperation without sanctions.