Online Dual-Resolution 3D Map Caching Algorithm Using MILP as a Neural Network Proxy

• Published in: Proceedings - International Conference on Computer Communications and Networks pp. 1 - 9
• Main Authors: Yang, Chun-An, Huang, Guo-Wei, Lo, Kuan-Hsiang, Sun, Shao-Lun, Chen, Yu-Wen, Kuo, Jian-Jhih, Tsai, Ming-Jer
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.08.2025
• Subjects: Costs; Deep reinforcement learning; information freshness; map scale; Mixed integer linear programming; online caching system; Optimization; optimization problem; Prediction algorithms; Roads; Safety; Simulation; Three-dimensional displays; Traffic control; vehicular network
• ISSN: 2637-9430
• DOI: 10.1109/ICCCN65249.2025.11133846

To provide real-time map information, roadside units (RSUs) near vehicles cache 3D map tiles to reduce transmission delays and alleviate backhaul congestion. Unfortunately, traditional cache problems primarily focus on maximizing the cache hit rate (i.e., popularity) while often neglecting other criteria, thus suppressing caching efficiency. For instance, the priority of each map tile may vary for different vehicles based on their route plans and current positions. Additionally, caching a greater number of small-scale map tiles with broader coverage helps serve more vehicles' requests and reduce cache misses. However, these tiles may only be sufficient for vehicles that can tolerate lower detail levels, as they provide less information than large-scale tiles. Furthermore, ensuring the freshness of cached map tiles is essential for maintaining driving safety. To address the interplay among popularity, priority, map scale, and information freshness in online map caching, this paper formulates an online optimization problem and proposes a novel online algorithm called ADAM, which integrates mixed-integer linear programming (MILP) with deep reinforcement learning. The ADAM predicts future total costs to support caching decisions by leveraging an ingeniously designed MILP that emulates the behavior of a neural network. Finally, simulation results manifest that the proposed ADAM outperforms existing methods by an average of 50%.