Generalization-Enhanced DRL-Based Resource Allocation in Wireless Communication Networks With Dynamic User Loads

Deep reinforcement learning (DRL)-based methods face retraining overhead when addressing resource allocation with time-varying action spaces caused by dynamic user loads. Leveraging the Markovian chain transmission characteristics of historical information, we propose a state decomposition approach that converts time-varying action spaces into adaptive adjustments of virtual state quantities. Based on this, a lightweight policy-action space approach is designed to solve the channel conflict problem. We prove the Markovianity of the decomposed virtual state and analyze the impact of Markovian/non-Markovian dynamic user load models on DRL performance. Simulations validate the adaptability of our algorithm in dynamic user load scenarios, which outperforms the distributed decision-making algorithm based on attention mechanism.