Deep Learning Meets Wireless Network Optimization: Identify Critical Links

Liu, Lu; Yin, Bo; Zhang, Shuai; Cao, Xianghui; Cheng, Yu

doi:10.1109/tnse.2018.2827997

Cited by 57 publications

(25 citation statements)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In [31], the authors considered solving a resource allocation problem for D2D communication using a Q-learning-based method. In [32], the authors applied deep learning to reduce the complexity of solving a wireless networks optimization problem. The authors in [33] proposed a resource allocation strategy using cooperative reinforcement learning.…”

Section: Related Workmentioning

confidence: 99%

A Deep Learning Approach for Maximum Activity Links in D2D Communications

Zhang

Palmieri

et al. 2019

Sensors

View full text Add to dashboard Cite

Mobile cellular communications are experiencing an exponential growth in traffic load on Long Term Evolution (LTE) eNode B (eNB) components. Such load can be significantly contained by directly sharing content among nearby users through device-to-device (D2D) communications, so that repeated downloads of the same data can be avoided as much as possible. Accordingly, for the purpose of improving the efficiency of content sharing and decreasing the load on the eNB, it is important to maximize the number of simultaneous D2D transmissions. Specially, maximizing the number of D2D links can not only improve spectrum and energy efficiency but can also reduce transmission delay. However, enabling maximum D2D links in a cellular network poses two major challenges. First, the interference between the D2D and cellular communications could critically affect their performance. Second, the minimum quality of service (QoS) requirement of cellular and D2D communication must be guaranteed. Therefore, a selection of active links is critical to gain the maximum number of D2D links. This can be formulated as a classical integer linear programming problem (link scheduling) that is known to be NP-hard. This paper proposes to obtain a set of network features via deep learning for solving this challenging problem. The idea is to optimize the D2D link schedule problem with a deep neural network (DNN). This makes a significant time reduction for delay-sensitive operations, since the computational overhead is mainly spent in the training process of the model. The simulation performed on a randomly generated link schedule problem showed that our algorithm is capable of finding satisfactory D2D link scheduling solutions by reducing computation time up to 90% without significantly affecting their accuracy.

show abstract

Section: Related Workmentioning

confidence: 99%

A Deep Learning Approach for Maximum Activity Links in D2D Communications

Zhang

Palmieri

et al. 2019

Sensors

View full text Add to dashboard Cite

show abstract

“…To circumvent this issue, a novel approach was proposed in [1], which approximates the optimized solution as a function of instantaneous channel gains with neural networks by leveraging the ability of universal approximation. Based on the observation that significant computing efforts are wasted due to repeatedly solving a problem under similar conditions, a deep learning framework was proposed in [5] to find the latent relationship between flow information and link usage by learning from past computation experience. To learn the optimal predictive resource allocation under the QoS constraint of video-on-demand service, a deep neural network (DNN) was designed in [6].…”

Section: Introductionmentioning

confidence: 99%

“…To learn the optimal predictive resource allocation under the QoS constraint of video-on-demand service, a deep neural network (DNN) was designed in [6]. By training the DNNs offline, an approximated solution can be obtained with low complexity online [1,5,6], say about 1% of the original numerical optimization [6]. Such an idea of "learning to optimize" can be regarded as a kind of computing offloading over time, which shifts the computations from online to offline.…”

Section: Introductionmentioning

confidence: 99%

“…The performance of supervised learning highly depends on data, especially labels. In [1,5], the approximated solution can achieve around 90% performance of the numerical solution. In [6], even with sophisticated DNN trained with 1,5000 labels generated from the global optimal solution, the QoS of each user can still not be satisfied.…”

Section: Introductionmentioning

confidence: 99%

“…In [6], even with sophisticated DNN trained with 1,5000 labels generated from the global optimal solution, the QoS of each user can still not be satisfied. For the application scenarios in [1,5,6] where the users can be served with best effort, such performance degradation is acceptable. For the applications with stringent QoS constraints such as ultra-reliable and low-latency communications (URLLC), is the DNN-based solutions still useful?…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Learning to Optimize with Unsupervised Learning: Training Deep Neural Networks for URLLC

Sun

Yang

2019

2019 IEEE 30th Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC)

View full text Add to dashboard Cite

Learning the optimized solution as a function of environmental parameters is effective in solving numerical optimization in real time for time-sensitive applications. Existing works of learning to optimize train deep neural networks (DNN) with labels, and the learnt solution are inaccurate, which cannot be employed to ensure the stringent quality of service. In this paper, we propose a framework to learn the latent function with unsupervised deep learning, where the property that the optimal solution should satisfy is used as the "supervision signal" implicitly. The framework is applicable to both functional and variable optimization problems with constraints. We take a variable optimization problem in ultra-reliable and low-latency communications as an example, which demonstrates that the ultra-high reliability can be supported by the DNN without supervision labels.

show abstract