PLATO enables MARL agents to handle unbounded task and agent changes
Pointer-network actor adapts to new tasks and agents without retraining —
Deep Dive
PLATO is a multi-agent reinforcement learning framework that handles both agent and task openness without artificial bounds. It uses a pointer-network-based actor to output distributions directly over current tasks, paired with a graph neural network critic. Trained with multi-agent proximal policy optimization under centralized training and decentralized execution, it achieves strong zero-shot generalization on the MOASEI wildfire suppression benchmark, outperforming padding, masking, and graph-based baselines.
Key Points
- Pointer-network actor adapts to variable task sets without masking or retraining, supporting unbounded action spaces.
- Graph neural network critic dynamically encodes changing agent-task interactions via a topology that shifts with composition.
- Outperforms padding, masking, and graph-based baselines on the MOASEI wildfire suppression benchmark, with strong zero-shot generalization.
Why It Matters
PLATO removes artificial limits on MARL deployments, making autonomous teams viable in unpredictable real-world environments like disaster response.