Dynamic resource management for network slicing: overcoming AI dimensionality via sliding window proxies

In 5 G/6 G network slicing, dynamic resource orchestration is often hindered by the dimensionality constraints of Deep Reinforcement Learning (DRL) algorithms. Models trained on fixed input/output tensor dimensions suffer from severe dimensionality mismatches when network topologies expand, forcing orchestrators to undergo computationally prohibitive retraining processes. Furthermore, limited generalization leads to policy degradation, causing inefficiencies such as suboptimal Modulation and Coding Scheme (MCS) matching and bandwidth fragmentation. To address these issues, this study proposes Monte Carlo Tree Searching with Sliding Window Convolution (MCTS-SWC). By deploying small, cost-effective unit-scale proxy models, MCTS-SWC utilizes a sliding window technique to scan and stitch local probability distributions across larger, untrained network matrices. Results show that MCTS-SWC delivers up to 95% of the optimal throughput within trained dimensions. For expanded network topologies, the proxy model reduces training time by 96.38%. While this massive operational gain introduces a performance trade-off—achieving 86% of optimal throughput in demanding scenarios compared to the 95% of standard models—it provides near-optimal efficiency under low-traffic conditions. Ultimately, within the range of network scales evaluated in this study (6 × 4 up to 30 × 26), MCTS-SWC facilitates a highly adaptable proxy architecture that operates independently of network scale variations, significantly mitigating the retraining overhead for dynamic network management.