PLATO Enables Multi-Agent Reinforcement Learning in Open Systems

Multi-agent reinforcement learning (MARL) traditionally struggles in "open agent systems" (OASYS), environments where the number of agents and tasks can unpredictably change. Current solutions often rely on restrictive assumptions or partial approaches like padding, which introduce artificial bounds. This new research presents PLATO (Pointer Learner for Agent and Task Openness), a framework specifically designed to address both agent and task openness without these limitations. PLATO integrates a pointer-network-based actor, which directly outputs distributions over the current task set, allowing for dynamic action spaces without requiring masking or retraining. It also features a centralized graph neural network (GNN) critic that models agent-task interactions, adapting its structure as agents and tasks evolve. Evaluated in a wildfire suppression domain, PLATO demonstrated strong performance and superior zero-shot generalization compared to state-of-the-art baselines, proving its effectiveness in truly open multi-agent environments.