Learning to cooperate in population games with scalable and resilient intrinsic rewards

Cooperation emerges when agents work with others to increase their collective welfare despite individual costs. It is an essential feature of social interactions among humans, but with humans increasingly delegating their tasks to AI, it is important that artificial agents can also learn to navigate the incentives of cooperation. This research contributes to developing learning agents capable of fostering cooperation in large populations. Using intrinsic reward functions we identify effective mechanisms for promoting and stabilising cooperation across different scenarios. This approach is promising for future applications of multi-agent systems, where decentralisation across many diverse agents will be prevalent.