TY - GEN
T1 - RL-Based Energy-Efficient Computation Offloading under Deadline Constraints in Multi-User Systems
AU - Regmi, Manish
AU - Ghasemi, Mehdi
AU - Vrudhula, Sarma
N1 - Publisher Copyright: © 2026 IEEE.
PY - 2026
Y1 - 2026
N2 - This paper addresses the challenge of energy-efficient computation offloading in multi-user, single-cloudlet systems under deadline constraints. While offloading AI workloads to a shared cloudlet can augment edge device capabilities, it faces cloudlet contention and wireless channel variability, complicating energy-latency trade-offs. We propose a fine-grained, level-wise reinforcement learning (RL) framework that makes per-level offloading decisions in real-time. Our RL approach, based on an actor-critic (A2C) agent, dynamically adapts to bandwidth fluctuations and cloudlet queue states while adhering to fractionally allocated deadlines. Our method reduces energy consumption by 33% compared to local-only execution, by 42% compared to a coarse-grained approach, and by 56% compared to all-cloudlet offloading, while consistently meeting deadlines and maintaining low runtime overhead.
AB - This paper addresses the challenge of energy-efficient computation offloading in multi-user, single-cloudlet systems under deadline constraints. While offloading AI workloads to a shared cloudlet can augment edge device capabilities, it faces cloudlet contention and wireless channel variability, complicating energy-latency trade-offs. We propose a fine-grained, level-wise reinforcement learning (RL) framework that makes per-level offloading decisions in real-time. Our RL approach, based on an actor-critic (A2C) agent, dynamically adapts to bandwidth fluctuations and cloudlet queue states while adhering to fractionally allocated deadlines. Our method reduces energy consumption by 33% compared to local-only execution, by 42% compared to a coarse-grained approach, and by 56% compared to all-cloudlet offloading, while consistently meeting deadlines and maintaining low runtime overhead.
KW - Computation offloading
KW - Edge computing
KW - reinforcement learning (RL)
UR - https://www.scopus.com/pages/publications/105045824762
UR - https://www.scopus.com/pages/publications/105045824762#tab=citedBy
U2 - 10.1109/ICCSPA69228.2026.11600865
DO - 10.1109/ICCSPA69228.2026.11600865
M3 - Conference contribution
T3 - Proceedings of the ICCSPA 2026 - 7th International Conference on Communications, Signal Processing, and their Applications
BT - Proceedings of the ICCSPA 2026 - 7th International Conference on Communications, Signal Processing, and their Applications
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 7th International Conference on Communications, Signal Processing, and their Applications, ICCSPA 2026
Y2 - 15 June 2026 through 18 June 2026
ER -