Tag: #multi-agent-rl7 results
KD-MARL introduces a two-stage framework for resource-aware knowledge distillation in multi-agent reinforcement learning, transferring coordinated behavior from centralized experts to lightweight decentralized students. It uses distilled advantage signals and structured supervision to preserve coordination without a critic, supporting heterogeneous architectures. Benchmarks show over 90% performance retention with up to 28.6x FLOPs reduction on SMAC and MPE.
This arXiv paper introduces the Efficiency Attenuation Phenomenon (EAP), where AI agents in MARL develop inscrutable protocols outperforming human-like symbolic languages by 50.5% in navigation tasks. It challenges the Language of Thought (LoT) hypothesis, arguing optimal cognition relies on sub-symbolic computations. The findings bridge AI, cognitive science, and philosophy with ethics implications.
Researchers demonstrate that sequence models' in-context learning induces cooperation in multi-agent RL without hardcoded co-player assumptions or timescale separation. Training against diverse co-players leads to best-response strategies on intra-episode timescales. This naturally emerges mutual shaping via extortion vulnerability, providing a scalable path to cooperative behaviors.
CAFE reformulates automated feature engineering as a causally-guided sequential decision process using causal discovery for soft priors and multi-agent RL for construction. It outperforms baselines by up to 7% on 15 benchmarks and reduces performance drops 4x under covariate shifts. The framework produces compact, stable features with reliable attributions.
This research introduces a hierarchical reinforcement learning framework that balances coordination efficiency with rigorous safety constraints. By utilizing a constraint manifold at the low level, the system ensures safety while maintaining stable training dynamics in multi-agent environments.
2-year project optimizes live line-haul logistics with MARL 'Fleet Manager' (PPO) for high-level clustering/dispatch and LP 'Dock Worker' for bin packing/TSP. Achieves zero-shot generalization via normalized density maps. Full architecture deep dive shared.
Co-jump enables two quadrupeds to synchronize jumps up to 1.5m via MAPPO and curriculum, without communication. Achieves 144% height gain over solo robots using proprioception. Transfers from sim to hardware.