Cooperation Under Uncertain Incentive Alignment: A Multi-Agent Reinforcement Learning Perspective

Cooperation is essential for addressig many complex challenges such as climate change mitigation, development of renewable energy policies, and urban traffic management. These goals require cooperation among agents, but traditional models for studying cooperation often fail to capture the complexity of real-world interactions.
To address this problem, this thesis introduces the Extended Public Goods Game (EPGG), a framework that allows for the investigation of cooperation emergence and optimal decision-making by addressing two aspects that are often overlooked: varying levels of incentive alignment and uncertainty regarding incentive alignment. Moreover, we employ decentralized multi-agent reinforcement learning (MARL) to study the behaviour of artificial agents trained within the EPGG framework.
We further explore three agent-dependent factors influencing cooperation: emergent communication, behavioural mechanisms, and risk preferences. We observe that emergent communication protocols can enhance cooperation in situations with symmetric information among agents, but enable deception under asymmetric information. Behavioural mechanisms such as reputation and intrinsic rewards sustain cooperation under uncertainty while preserving competition when needed. Lastly, risk preferences reshape the landscape of game equilibria. Experimental results show that risk-seeking agents cooperate more, while risk-averse ones tend to defect, with these tendencies being amplified under uncertainty.
By considering incentive dynamics, uncertainty, and agent-specific behaviours, this work provides a more nuanced understanding of cooperation and its emergence in systems of multiple agents, contributing to the development of artificial agents better equipped to navigate real-world scenarios.