PLATO Enhances Multi-Agent Reinforcement Learning for Open Systems.

Open agent systems (OASYS), where the number and types of agents and tasks can change unpredictably, present a significant challenge for traditional multi-agent reinforcement learning (MARL). Existing MARL methods typically assume fixed state and action spaces, often relying on padding or masking, which introduce artificial limitations. PLATO (Pointer Learner for Agent and Task Openness) is introduced as a solution to this fundamental problem. PLATO employs a pointer-network-based actor that directly outputs distributions over the current task set, allowing for dynamic action spaces without requiring retraining or masking. Complementing this, a centralized graph neural network (GNN) critic encodes agent-task interactions, adapting its structure as agents and tasks evolve. This architecture, trained with multi-agent proximal policy optimization, enables PLATO to handle both agent openness and task openness without the boundedness constraints of previous approaches. Evaluated in a wildfire suppression domain, PLATO demonstrated strong performance and consistent zero-shot generalization, proving its adaptability in highly dynamic, open multi-agent environments.