PLATO: A Pointer-Network Approach to Open Multi-Agent Systems
A new paper introduces PLATO (Pointer Learner for Agent and Task Openness), a multi-agent reinforcement learning framework designed for open agent systems where the sets of agents and tasks change unpredictably. Existing methods like padding and masking impose artificial bounds, while graph-based approaches handle only one dimension of openness. PLATO combines a pointer-network-based actor with a centralized graph neural network critic, trained via multi-agent proximal policy optimization under centralized training and decentralized execution. The pointer actor outputs distributions directly over current tasks, enabling adaptation to varying agent and task sets without fixed state or action spaces. The paper is available on arXiv under ID 2607.25082.
Key facts
- PLATO stands for Pointer Learner for Agent and Task Openness.
- It addresses agent openness and task openness in multi-agent reinforcement learning.
- The framework uses a pointer-network-based actor and a centralized GNN critic.
- Training uses multi-agent proximal policy optimization.
- It operates under centralized training and decentralized execution.
- Existing methods like padding and masking impose artificial bounds.
- Graph-based methods handle only one dimension of openness.
- The paper is published on arXiv with ID 2607.25082.
Entities
Institutions
- arXiv