Modern connected vehicles (CVs) frequently require diverse types of content for mission-critical decision-making and onboard users' entertainment. These contents are required to be fully delivered to the requester CVs within stringent deadlines that the existing radio access technology (RAT) solutions may fail to ensure. Motivated by the above consideration, this paper exploits content caching in vehicular edge networks (VENs) with a software-defined user-centric virtual cell (VC) based RAT solution for delivering the requested contents from a proximity edge server. Moreover, to capture the heterogeneous demands of the CVs, we introduce a preference-popularity tradeoff in their content request model. To that end, we formulate a joint optimization problem for content placement, CV scheduling, VC configuration, VC-CV association and radio resource allocation to minimize long-term content delivery delay. However, the joint problem is highly complex and cannot be solved efficiently in polynomial time. As such, we decompose the original problem into a cache placement problem and a content delivery delay minimization problem given the cache placement policy. We use deep reinforcement learning (DRL) as a learning solution for the first sub-problem. Furthermore, we transform the delay minimization problem into a priority-based weighted sum rate (WSR) maximization problem, which is solved leveraging maximum bipartite matching (MWBM) and a simple linear search algorithm. Our extensive simulation results demonstrate the effectiveness of the proposed method compared to existing baselines in terms of cache hit ratio (CHR), deadline violation and content delivery delay.

翻译：现代互联车辆（CV）在执行关键任务决策和车载用户娱乐时，频繁需要不同类型的内容。这些内容必须在严格的截止时间内完整传输至请求车辆，而现有无线接入技术（RAT）方案可能无法确保这一点。基于上述考虑，本文在车载边缘网络（VEN）中利用内容缓存技术，结合基于软件定义、面向用户的虚拟小区（VC）RAT方案，从邻近边缘服务器交付所请求的内容。此外，为捕获互联车辆异构需求特征，我们在其内容请求模型中引入偏好-流行度权衡。据此，我们构建了内容放置、互联车辆调度、虚拟小区配置、虚拟小区-互联车辆关联及无线资源分配的联合优化问题，以最小化长期内容投递延迟。然而，该联合问题高度复杂，无法在多项式时间内高效求解。因此，我们将原始问题分解为缓存放置子问题，以及在给定缓存放置策略下的内容投递延迟最小化子问题。针对第一子问题，采用深度强化学习（DRL）作为学习方案。进一步，将延迟最小化问题转化为基于优先级的加权和速率（WSR）最大化问题，并借助最大二分图匹配（MWBM）及简单线性搜索算法求解。广泛仿真结果表明，在缓存命中率（CHR）、截止时间违反率和内容投递延迟方面，所提方法相较于现有基线方案具有显著有效性。