DDPG-based Wireless Resource Allocation for Time-Constrained Applications

This paper presents a novel model-free resource allocation framework for the downlink of 5G cellular networks to guarantee stringent QoS requirements in wireless applications. A Deep deterministic policy gradient (DDPG) agent with a modified Genetic Algorithm (GA) based resource allocation framework is proposed to balance the tradeoffs between reliability, latency, and data rate. Any feasible point in the rate-latency-reliability domain can be achieved with this approach. Compared to state-of-the-art approaches DDPG-Dual and DDPG-PSO, the proposed model achieves higher reliability and scalability in joint optimization with QoS constraints. Specifically, the proposed model guarantees the expected reliability with 25% and 42.86% improvement respectively over the compared models. In terms of conventional effective bandwidth approach, the proposed model achieves 30.82% improvement of energy efficiency under the same QoS constraints. Moreover, the proposed model offers a practical solution, namely, three times faster convergence and only 6.7% of the scheduling time compared to the ground truth Dual decomposition optimization.