Handovers (HOs) have been envisioned to be more challenging in 5G networks due to the inclusion of millimetre wave (mm-wave) frequencies, resulting in more intense base station (BS) deployments. This, by its turn, increases the number of HOs taken due to smaller footprints of mm-wave BSs thereby making HO management a more crucial task as reduced quality of service (QoS) and quality of experience (QoE) along with higher signalling overhead are more likely with the growing number of HOs. In this paper, we propose an offline scheme based on double deep reinforcement learning (DDRL) to minimize the frequency of HOs in mm-wave networks, which subsequently mitigates the adverse QoS. Due to continuous and substantial state spaces arising from the inherent characteristics of the considered 5G environment, DDRL is preferred over conventional $Q$-learning algorithm. Furthermore, in order to alleviate the negative impacts of online learning policies in terms of computational costs, an offline learning framework is adopted in this study, a known trajectory is considered in a simulation environment while ray-tracing is used to estimate channel characteristics. The number of HO occurrence during the trajectory and the system throughput are taken as performance metrics. The results obtained reveal that the proposed method largely outperform conventional and other artificial intelligence (AI)-based models.

由于包含毫米波（mm-wave）频率，导致5G网络中的切换（HOs）更具挑战性，导致基站（BS）部署更加密集。进而，由于毫米波BS的占用空间较小，因此增加了HO的数量，从而使HO管理成为一项更关键的任务，因为降低了服务质量（QoS）和体验质量（QoE）以及更高的信令开销随着HO数量的增加，可能性更大。在本文中，我们提出了一种基于双深度强化学习（DDRL）的离线方案，以最小化毫米波网络中HO的频率，从而减轻不利的QoS。由于考虑到的5G环境的固有特性会产生连续且大量的状态空间，因此DDRL优于传统的DDRL$问$学习算法。此外，为了减轻在线学习策略在计算成本方面的负面影响，本研究采用了离线学习框架，在模拟环境中考虑了已知轨迹，而使用光线跟踪来估计信道特性。轨迹中HO发生的次数和系统吞吐量被视为性能指标。获得的结果表明，所提出的方法大大优于传统的和其他基于人工智能（AI）的模型。