Decomposition Model Assisted Energy-Saving Design in Radio Access Network

TL;DR

This work tackles the challenge of high energy consumption in ultra-dense radio access networks by introducing a decomposition model that uses a softgoal interdependency graph to break the network-level energy-saving objective into executable, multi-granularity operations. The run-time decision making is guided by a deep Q-network that leverages action spaces defined by the decomposition, with weights and conflicts between objectives identified to prune suboptimal choices. The authors demonstrate that coupling the decomposition model with DQN accelerates training and improves trade-offs among energy, throughput, and first-packet delay, validated through simulation on multi-BS/multi-UE scenarios. The approach offers a practical framework for real-time, multi-objective energy management in 6G RANs and highlights the value of integrating decomposition with reinforcement learning for network-wide optimization.

Abstract

The continuous emergence of novel services and massive connections involve huge energy consumption towards ultra-dense radio access networks. Moreover, there exist much more number of controllable parameters that can be adjusted to reduce the energy consumption from a network-wide perspective. However, a network-level energy-saving intent usually contains multiple network objectives and constraints. Therefore, it is critical to decompose a network-level energy-saving intent into multiple levels of configurated operations from a top-down refinement perspective. In this work, we utilize a softgoal interdependency graph decomposition model to assist energy-saving scheme design. Meanwhile, we propose an energy-saving approach based on deep Q-network, which achieve a better trade-off among the energy consumption, the throughput, and the first packet delay. In addition, we illustrate how the decomposition model can assist in making energy-saving decisions. Evaluation results demonstrate the performance gain of the proposed scheme in accelerating the model training process.

Figures (7)

• Figure 1: 5G radio access networks scenario.
• Figure 2: The specific implementation flow of the decomposition model assisted energy-saving scheme.
• Figure 3: The combinations of the three energy-saving operations at different granularities are stored as JSON format.
• Figure 4: The adaptive adjustment time testing of the decomposition model assisted scheme.
• Figure 5: The effect of the decomposition model on DQN loss function.