<ul><li>This paper investigates the impact of relational state abstraction on sample efficiency and performance in collaborative Multi-Agent Reinforcement Learning (MARL).</li><li>The proposed abstraction is based on spatial relationships, leveraging spatial reasoning in real-world multi-agent scenarios.</li><li>The authors introduce MARC (Multi-Agent Relational Critic), a critic architecture that incorporates spatial relational inductive biases by transforming the state into a spatial graph and processing it through a relational graph neural network.</li><li>Empirical analysis shows that MARC outperforms state-of-the-art MARL baselines in terms of sample efficiency, asymptotic performance, and potential for generalization, without requiring complex designs or task-specific engineering.</li></ul>

Investigating Relational State Abstraction in Collaborative MARL

Discover more