Transformer-Driven Distributed Reinforcement Learning for Efficient IoT Application Scheduling in Edge-Cloud Systems
Journal of Electronics and Power Engineering · 30.04.2026 · 3 (1) · 1-8
DOI: 10.69996/jepe.2026001 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Electronics and Power Engineering, online erschienen am 30.04.2026. Verfasst von Padala Sravan und 4 weiteren Autor:innen.
Abstract (Originalsprache)
The high rate of Internet of Things (IoT) applications development and their rampant expansion has brought unprecedented complexities to the modern network inferences such that effective scheduling of tasks in the context of both hybrid edge and cloud computing is a critical as well as a highly difficult issue. The dynamism, heterogeneity of resources and stochastic nature of these distributed environments demand highly adaptive and intelligent solutions in real time resource management. Whereas current methods of centralized deep learning have been scaled to these requirements, they are usually severely disadvantageous such as training times are prohibitive, convergence is slow, exploration is too expensive, and they simply cannot easily handle the complex, interdependent task structures that characterize modern IoT processes. To successfully eliminate all these limitations, the study proposes a new flexible and highly scalable framework that is based on the Distributed Reinforcement Learning, dubbed TF-DDRL, the framework that is specially created and designed to be optimal in terms of scheduling heterogeneous applications that will be run by the IoT. With a strong basis of Actor-Critic, TF-DDRL is able to allocate the learning process to several servers and uses a strict off-policy correction strategy to stabilize training. Importantly, the proposed architecture is innovative in the way that it combines Prioritized Experience Replay (PER) to reduce exploration costs significantly, speed up the learning phase, and, at the same time, make use of complex Transformer models to precisely model and map out longterm dependencies on tasks throughout the network. Through a prudent tradeoff between existing conflicting goals, large-scale experimental studies provide conclusive evidence that the TFDDRL framework can generate radical improvements in various systemlevel metrics including reducing response time by 60%t, energy use by 51%, monetary costs by 56%, and the overall weighted cost by a full 58%, thereby providing an incredibly efficient and sustainable framework of next-generation IoT systems.
Zitieren
APA
Sravan, P., Swetha, B., Bharadwaj, S. S., Chandhana, S. S., & Salam, M. A. (2026). Transformer-Driven Distributed Reinforcement Learning for Efficient IoT Application Scheduling in Edge-Cloud Systems. Journal of Electronics and Power Engineering, 3(1), 1-8. https://doi.org/10.69996/jepe.2026001
Vancouver
Sravan P, Swetha B, Bharadwaj SS, Chandhana SS, Salam MA. Transformer-Driven Distributed Reinforcement Learning for Efficient IoT Application Scheduling in Edge-Cloud Systems. Journal of Electronics and Power Engineering. 2026;3(1):1-8. doi:10.69996/jepe.2026001
BibTeX
@article{Sravan2026,
author = {Sravan, Padala and Swetha, Bommadeni and Bharadwaj, Sangaras Sai and Chandhana, Samudrala Siri and Salam, Md Abdul},
title = {Transformer-Driven Distributed Reinforcement Learning for Efficient IoT Application Scheduling in Edge-Cloud Systems},
journal = {Journal of Electronics and Power Engineering},
year = {2026},
volume = {3},
number = {1},
pages = {1--8},
doi = {10.69996/jepe.2026001}
}
EUVIMED