Research map / Multi-agent systems and games / Cooperative multi-agent RL

PLATO: Pointer Learner for Agent and Task Openness

NeurIPS 2026 · Multi-agent systems and games › Cooperative multi-agent RL · approach: MARL benchmarks and grouping · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PLATO, a pointer-network actor with a bipartite GNN critic trained via MAPPO, to handle both agent and task openness in multi-agent RL without padding or masking, evaluated on the MOASEI wildfire domain.

Problem

Multi-agent RL assumes fixed agent and task sets, but open agent systems have agents and tasks entering and leaving unpredictably; existing methods handle task or agent openness only partially or rely on padding and bounds.

Approach

A content-based pointer actor outputs a categorical distribution over the currently available tasks (plus no-op) using additive attention over task keys and a permutation-invariant team query; a centralized bipartite agent-task GCN critic evaluates the open state. Formalized as a TaAgO-MG extending MOHITO's generator to agent sets.

Key results

• On 3x3 grid, PLATO achieves 112.83±12.62 return in S2 vs 59.75 for DICG and 42.36 for MOHITO, statistically significant.
• On 4x4 grid S3, PLATO reaches 149.35±27.33 return vs 74.39 for DICG.
• Zero-shot trained on 3x3 and tested on 5x5 in S2, PLATO gets 80.79±21.55 vs 41.51 for DICG.

Surprising findings

• MOHITO collapses to negative returns in S2 at 4x4 under zero-shot transfer from 2x3.
• In native evaluation, LSTM encoder and dot-product scorer show no significant advantage; benefits appear only in zero-shot S3.

Limitations

• Assumes one action per task and full observability of agents and tasks.
• Evaluated only in the wildfire domain on small grids (up to 5x5).
• Efficiency results are mixed; Weakest leads reward per fight in S1 on 3x3.

Experimental setup

Datasets: MOASEI wildfire domain (simulated)

Benchmarks: MOASEI wildfire suppression (2x3, 3x3, 4x4, 5x5 grids; setups S0-S3)

Metrics: episode return reward per fight action zero-shot episode return

Baselines: MOHITO DICG DGN NOOP Random Strongest Weakest

Builds on: MAPPO pointer networks MOHITO DGN DICG

Closest papers by meaning

• Learning Agentic Policy from Action Guidance (NeurIPS 2026 · similarity 0.87)
• Group-Graph Policy Optimization for Long-Horizon Agentic Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
• AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.86)
• Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use (ICML 2026 · similarity 0.86)
• Pessimistic Latent Task-aware Optimization for Robust Offline Meta-Reinforcement Learning (NeurIPS 2026 · similarity 0.86)
• Agentick: A Unified Benchmark for General Sequential Decision-Making Agents (NeurIPS 2026 · similarity 0.86)