The «Stag Hunt» Social Dilemma in Multi-Agent Reinforcement Learning

This paper discusses the identification and resolution of social dilemmas in multi-agent reinforcement learning. A multi-agent environment was developed to simulate the movement of autonomous cars on a circular two-lane road, with two agent cars independently trained by a neural network. An independent Q-learning algorithm (IQN) was used for effective environment learning, allowing each agent to find its own strategy to improve learning efficiency. During the agents' training process, the «stag hunt» social dilemma was identified, which arose from agents refusing to cooperate with each other due to fear of betrayal. Experimental results showed that the IQN algorithm effectively resolves social dilemmas in various multi-agent scenarios.