What question did this study set out to answer?

The aim is to improve task offloading and resource allocation efficiency in mobile edge computing using deep reinforcement learning.

March 10, 2026Open Access

EdgeAutoOffloadAI a deep reinforcement learning framework for adaptive task offloading and resource management in mobile edge computing

Key Points

The aim is to improve task offloading and resource allocation efficiency in mobile edge computing using deep reinforcement learning.
Developed a deep reinforcement learning framework called EdgeAutoOffloadAI.
Adopted Proximal Policy Optimisation (PPO) with a Transformer-based encoder for training.
Implement a reward function using Jain’s fairness index and Conditional Value-at-Risk (CVaR).
Conducted extensive simulations using NS-3.
Reduced average latency by up to 25%.
Missed deadlines 30% less often.
Increased resource fairness by 15% compared to baseline methods.

Abstract

As emerging low-latency, compute-intensive applications, e.g., autonomous driving, industrial IoT, and augmented reality, experience explosive growth, this demand has driven a surge in real-time computing services. Dealing with such demands is not feasible in traditional cloud computing architectures due to high communication latency and heavy network congestion, leading to a natural performance decrease for mobile users. Accordingly, the Mobile Edge Computing (MEC) concept has emerged as a new approach to push resources to edge devices. However, task offloading and resource allocation in MEC systems become challenging due to the dynamic nature of network conditions, the variety of device capabilities, and the unpredictability of end-user behaviour. Existing heuristic and optimisation approaches are unavoidably inflexible. At the same time, early DRL methods such as Deep Q-Networks (DQN) only consider discrete action spaces and lack efficient scalability, which makes them very ill-suited for continuous resource provisioning. To relax these limitations, we present EdgeAutoOffloadAI. This novel DRL framework adopts Proximal Policy Optimisation (PPO) with a Transformer-based encoder for efficient training of adaptive task offloading and resource allocation. Based on the Transformer design, this abstraction enables the network to model intricate relationships between users, edge servers, and network states while remaining scalable across diverse network sizes. The reward function is also improved with Jain’s fairness index for resource fairness and Conditional Value-at-Risk (CVaR) for handling extreme latency events, which enhance the algorithm’s robustness and fairness compared to extensive NS-3 simulations. We demonstrate that EdgeAutoOffloadAI reduces the average latency by up to 25%, misses deadlines 30% less often, and increases fairness by 15% compared with state-of-the-art baseline methods. These results highlight the practical suitability of EdgeAutoOffloadAI for real-world MEC applications that require scalability, reliability, and effective resource utilisation in future-generation edge computing systems.

AI से पूछें

Bookmark

View Full Paper