Research map / Multi-agent systems and games / Cooperative multi-agent RL
A Separation Principle for Cooperative Multi-Agent Reinforcement Learning
Lucia Pezzetti, Nicolas Lanzetti, Antonio Terpin, Florian Dorfler, Giorgia Ramponi
Read the paperCheck your idea against 137,217 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.
Studies cooperative MARL with decoupled noisy agent dynamics coupled through a population objective, motivated by fleet mobility, and proposes a separation principle for learning in this setting.
Problem
Cooperative multi-agent RL where agents evolve under decoupled noisy dynamics but are coupled via a population-level objective (e.g., a mobility operator managing a vehicle fleet to meet customer demand); existing MARL algorithms are said to not exploit this structure.
Approach
Exploits the decoupled-dynamics-plus-population-objective structure via a separation principle, as stated in the title; the abstract is truncated so the specific mechanism is not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| separate assignment (optimal transport) from single-agent routing instead of learning a joint MARL policy | SALT · Stochastic grid world (N=7) | average travel time: 23-61% reduction vs MARL baselines | Improved |
| swap the single-agent learner among Q-learning, DDQN, PPO and A2C (ablation) | SALT · Stochastic grid world (N=7) | average travel time | No clear effect |
| use target-conditioned single-agent PPO in SALT instead of MAPPO as population size grows | SALT-PPO · Grid world scaling, N=2 to 128 | environment interactions to reach-rate milestones: up to 480x fewer interactions | Improved |
| scale population to N=8 and beyond with MAPPO versus SALT-PPO | MAPPO · Grid world scaling, 2e9 interaction budget | reach-rate milestone: MAPPO fails 70% milestone at N=8; SALT-PPO reaches 99% up to N=128 | Hurt |
| learned target-conditioned Q policy with optimal transport assignment versus static shortest-path routing | SALT-Q · Southern Manhattan road network | travel time improvement: 14.8% (N=1) to 21.3% (N=100) | Improved |
| evaluate policy trained with 20 vehicles on unseen fleet sizes without retraining | SALT-Q · Southern Manhattan road network, fleets of 1-100 | travel time improvement | No clear effect |
Closest papers by meaning
- Learning Joint Behaviors with Large Variations (AAAI 2025 · similarity 0.89)
- HyperMARL: Adaptive Hypernetworks for Multi-Agent RL (NeurIPS 2025 · similarity 0.88)
- One Step is Enough: Multi-Agent Reinforcement Learning Based on One-Step Policy Optimization for Order Dispatch on Ride-Sharing Platforms (NeurIPS 2026 · similarity 0.88)
- Foundations of Multi-Agent Learning in Dynamic Environments: Where Reinforcement Learning Meets Strategic Decision-Making (AAAI 2025 · similarity 0.87)
- Mean-Field Sampling for Cooperative Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Multi-agent cooperation through learning-aware policy gradients (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
