Research map / Multi-agent systems and games / Human-AI and goal-conditioned
MAT-Agent: Adaptive Multi-Agent Training Optimization
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 MAT-Agent, a multi-agent DQN-based framework that dynamically selects augmentation, optimizer, learning-rate schedule and loss per training step for multi-label image classification, reporting gains over PAT-T and HSQ-CvN on Pascal VOC, COCO and VG-256.
Problem
Static training configurations for multi-label image classification fail to adapt to evolving training dynamics; exhaustive hyperparameter/AutoML search is costly.
Approach
Treat each training component (augmentation, optimizer, LR schedule, loss) as an RL agent with a DQN policy, sharing a global composite reward, so the training configuration is adapted online per step rather than fixed in advance.
Key results
- On Pascal VOC, MAT-Agent reports mAP 97.4 (vs. 96.2 for PAT-T), OF1 92.3, CF1 91.4.
- On COCO, reports mAP 92.8 (vs. 92.0 for HSQ-CvN) and reaches 63.8% mAP in 47 epochs vs. about 80 for standard training (claimed 47% time reduction).
- On VG-256, reports mAP 60.9, OF1 70.8, CF1 61.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive multi-agent training optimization instead of the strongest listed baseline | MLIC models · Pascal VOC | mAP: 96.4 -> 97.4 | Helped |
| Use adaptive multi-agent training optimization instead of PAT-T | MLIC models · Pascal VOC | OF1: 91.1 -> 92.3 | Helped |
| Use adaptive multi-agent training optimization instead of PAT-T | MLIC models · Pascal VOC | CF1: 90.6 -> 91.4 | Helped |
| Use adaptive multi-agent training optimization instead of HSQ-CvN | MLIC models · COCO | mAP: 92.0 -> 92.8 | Helped |
| Use adaptive multi-agent training optimization instead of PAT-T | MLIC models · COCO | OF1: 87.6 -> 88.2 | Helped |
| Use adaptive multi-agent training optimization instead of HSQ-CvN | MLIC models · COCO | CF1: 86.6 -> 87.1 | Helped |
| Use adaptive multi-agent training optimization instead of PAT-T | MLIC models · VG-256 | mAP: 59.5 -> 60.9 | Helped |
| Use adaptive multi-agent training optimization instead of PAT-T | MLIC models · VG-256 | OF1: 69.8 -> 70.8 | Helped |
| Use adaptive multi-agent training optimization instead of PAT-T | MLIC models · VG-256 | CF1: 59.7 -> 61.1 | Helped |
| Train with adaptive multi-agent policies rather than DARTS | ResNet-101 trained on MS-COCO · MS-COCO → VOC | mAP: 73.8 -> 76.2 | Helped |
| Train with adaptive multi-agent policies rather than DARTS | ResNet-101 trained on MS-COCO · MS-COCO → NUS-WIDE | mAP: +2.8 points | Helped |
| Train with adaptive multi-agent policies rather than DARTS | ResNet-101 trained on MS-COCO · MS-COCO → OpenImages | mAP: +2.6 points | Helped |
Surprising findings
- Supplementary text states the 'structured communication channels and attention mechanisms' in the conclusion are future vision; no explicit inter-agent communication is implemented.
- Reported training-time reduction figures are inconsistent across the paper (47% vs 41.25%) and the sensitivity analysis text conflicts with its figure.
Limitations
- Per-epoch overhead of about 10% from four DQN agents.
- Many key details (action spaces, state features, reward weights) deferred to supplementary material; main text lacks them.
- Baseline numbers and comparison setup (all trained with same backbone) are not clearly specified.
Experimental setup
Datasets: Pascal VOC MS-COCO Visual Genome (VG-256) NUS-WIDE OpenImages
Benchmarks: Pascal VOC MS-COCO VG-256 NUS-WIDE OpenImages
Metrics: mAP OF1 CF1 Rare-F1
Baselines: ML-GCN C-Tran BalanceMix ASL ML-Decoder MLBOTE HSQ-CvN PAT-T PBT BOHB DARTS
Builds on: DQN QMIX VDN AutoAugment PBT
Closest papers by meaning
- CoPEFT: Fast Adaptation Framework for Multi-Agent Collaborative Perception with Parameter-Efficient Fine-Tuning (AAAI 2025 · similarity 0.84)
- GOAL: A Generalist Combinatorial Optimization Agent Learner (ICLR 2025 · similarity 0.84)
- ML-Agent: Reinforcing LLM Agents for Autonomous Machine Learning Engineering (ICML 2026 · similarity 0.84)
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.83)
- Adaptively Coordinating with Novel Partners via Learned Latent Strategies (NeurIPS 2025 · similarity 0.83)
- Towards Efficient Collaboration via Graph Modeling in Reinforcement Learning (AAAI 2025 · similarity 0.83)
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.