Buffer transference strategy for power control in B5G-ultra-dense wireless cellular networks

Beyond five generation (B5G) systems will demand strict and heterogeneous service requirements for the emerging applications. One solution to meet these demands is the dense deployment of small base stations to provide more capacity and coverage. However, this will lead to high power consumption and greenhouse emissions. Therefore, the resource control policies need to adapt to these network fluctuations to balance the power consumption and meet these demanding requirements. One approach is to implement intelligent algorithms for resource management, such as deep reinforcement learning models. These models can adapt to network changes and unknown conditions. However, while these models adjust to the new requirements, the performance is degraded due to state-space exploration. Therefore, accelerating the learning process is needed to minimize this performance degradation in dynamic environments. One of the approaches to address the above is to transfer the knowledge of other models to improve the learning process. This paper implements a training strategy in an ultra-dense network for power control. The method consists of reusing the previous experiences of models to train new models in complex environments, such as environments with more agents. We evaluate our proposal via simulation. The numerical results demonstrate that adding experiences to the buffer can accelerate the decision on power allocation to increase the network’s performance.