Accelerating Swarms : Harnessing Hardware Acceleration and Parallelization in Multi-Agent Reinforcement Learning - Archive ouverte HAL
Communication Dans Un Congrès Année : 2024

Accelerating Swarms : Harnessing Hardware Acceleration and Parallelization in Multi-Agent Reinforcement Learning

Résumé

In Multi-Agent Reinforcement Learning, researchers often face a challenging trade-off: the use of complex environments that demand substantial computational resources, or simpler dynamics for expedited execution, albeit at the cost of transferability to more realistic tasks. This article delves into the potential of vectorizable environments, which enable parallel environment rollout and fully harness the parallelization capabilities of modern GPUs. We present a comparison of various RL environment libraries, highlighting their features and limitations for end-to-end hardware-accelerated training pipelines. We observe that the most commonly used RL algorithm libraries have yet to fully embrace end-to-end hardware-accelerated training pipeline and the limited cross-compatibility between the frameworks used for hardware acceleration and parallelization in machine learning: PyTorch, TensorFlow, and JAX, limits the mix-and-match options for RL environments and algorithms libraries.
Fichier principal
Vignette du fichier
9.pdf (198.56 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04728894 , version 1 (09-10-2024)

Licence

Identifiants

  • HAL Id : hal-04728894 , version 1

Citer

Timothée Gavin, Simon Lacroix, Murat Bronz. Accelerating Swarms : Harnessing Hardware Acceleration and Parallelization in Multi-Agent Reinforcement Learning. annual International Micro Air Vehicle Conference and Competition, Sep 2024, Bristol, United Kingdom. pp.80-87. ⟨hal-04728894⟩
0 Consultations
0 Téléchargements

Partager

More