An adaptive deep reinforcement learning framework enhanced by broad reinforcement learning–based state transition for efficient task offloading and resource allocation at the network edge

Abstract Mobile Edge Computing (MEC) empowers resource-constrained devices to offload computationally intensive tasks to nearby edge servers, reducing latency and energy consumption. Existing deep reinforcement learning approaches have improved offloading and resource allocation decisions; however, their performance often depends on heuristic or manually tuned hyperparameters, which limit adaptability in non-stationary environments. To address this issue, we propose a Broad Reinforcement Learning–enhanced State-Driven Task Offloading Algorithm (BRL-STOA), which integrates fitness landscape analysis–based state representation, Broad Learning System (BLS)–assisted Q-value approximation, and Gaussian Mixture Model (GMM)–based data augmentation to allow efficient and adaptive decision-making. Unlike conventional deep reinforcement learning approaches that incur high computational overhead, the proposed framework employs a lightweight BLS structure to accelerate convergence while maintaining decision accuracy. Simulation results demonstrate that BRL-STOA achieves up to 21.2% reduction in execution cost, 38.8% lower energy consumption, and 9.1% improvement in system throughput compared with state-of-the-art meta-heuristic algorithms. Moreover, the proposed method increases the accumulated reward by approximately 8–14% across different state-space sizes and converges faster while maintaining improved stability.