This paper introduces RAMAC, a novel framework that extends the theory of realizable abstractions to cooperative multi-agent systems, addressing challenges like partial observability, coordination, and scalability. RAMAC decomposes joint decision-making into hierarchical levels, with high-level coordination policies manipulating abstract state representations and low-level agents executing temporally-extended options.
Key findings
RAMAC extends realizable abstractions theory to multi-agent settings, providing formal guarantees for policy quality.
The framework includes mechanisms for automatic abstraction discovery, eliminating manual hierarchy specification.
RAMAC integrates graph neural networks for inter-agent communication and option-critic methods for temporal abstraction.
The framework is evaluated on SMAC and cooperative navigation tasks, showing improved sample efficiency and coordination.
Limitations & open questions
The paper does not discuss the computational complexity of the proposed framework.
Evaluation is limited to specific benchmarks; broader applicability is yet to be demonstrated.