Deep reinforcement learning can promote sustainable human behaviour in a common-pool resource problem

Area of Science:

• Behavioral Economics
• Artificial Intelligence
• Game Theory

Background:

• Social dilemmas, such as resource allocation, often involve a conflict between individual gain and collective well-being.
• Effective resource allocation mechanisms are crucial for sustaining shared resources and promoting cooperation.
• Previous mechanisms often fail to adapt to dynamic player behavior and resource availability.

Purpose of the Study:

• To design an artificial intelligence (AI) based social planner using deep reinforcement learning (RL) to promote sustainable contributions in a multiplayer trust game.
• To create a simulated economy of human-like players to test and refine the RL mechanism.
• To maximize aggregate returns for all players while fostering fairness and cooperation.

Main Methods:

• Trained neural networks to simulate human player behavior in an iterated multiplayer trust game.
• Employed deep reinforcement learning (RL) to train a social planner mechanism.
• Developed a redistributive policy that conditions generosity on resource availability and incorporates temporary sanctions for defectors.

Main Results:

• The RL-designed mechanism significantly increased aggregate returns compared to baseline mechanisms.
• The mechanism achieved a more equal distribution of the surplus among players.
• The AI's policy, when translated into an explainable mechanism, was more favorably received by human participants.

Conclusions:

• Deep reinforcement learning can effectively design adaptive mechanisms for social dilemmas, promoting sustainable cooperation.
• AI-driven resource allocation can lead to both increased efficiency and greater equity in shared resource management.
• Explainable AI-derived policies enhance player acceptance and participation in cooperative systems.