Review
Neural networks in wireless networks: Techniques, applications and guidelines

https://doi.org/10.1016/j.jnca.2016.04.006Get rights and content

Abstract

The design of modern wireless networks, which involves decision making and parameter optimization, is quite challenging due to the highly dynamic, and often unknown, environmental conditions that characterize wireless networks. There is a common trend in modern networks to incorporate artificial intelligence (AI) techniques to cope with this design complexity. While a number of AI techniques have been profitably employed in the wireless networks community, the well-established AI framework of neural networks (NNs), well known for their remarkable generality and versatility, has been applied in a wide variety of settings in wireless networks. In particular, NNs are especially popular for tasks involving classification, learning, or optimization. In this paper, we provide both an exposition of common NN models and a comprehensive survey of the applications of NNs in wireless networks. We also identify pitfalls and challenges of implementing NNs especially when we consider alternative AI models and techniques. While various surveys on NNs exist in the literature, our paper is the first paper, to the best of our knowledge, which focuses on the applications of NNs in wireless networks.

Introduction

While computers are typically utilized in a problem solving capacity, traditionally this capacity has been restricted to the cases where the solution can be explicitly programmed. There are many problems in wireless networks where general programming techniques do not provide optimal or generalizable solutions. For example, it is very difficult to write programs that could allow cognitive radios to learn and adapt, or predict user mobility to optimize wireless network resources. Advances in artificial intelligence (AI) and machine learning (ML) allow computers to overcome this restrictive constraint by enabling inference and complex decision making (attributes that are typically associated with human intelligence). In particular, computer systems can use machine learning algorithms and techniques to learn from data allowing these systems to classify and predict (Mitchell, 1997, Mitchell, 2006). Broadly speaking, ML involves detecting patterns and learning correlations between input and output from past data. This knowledge is then used to extrapolate outputs, or infer, on future input data. ML is a vast field encompassing many different kinds of systems which themselves use many different models.

Artificial Neural Networks (ANNs), or simply Neural Networks (NNs) (Haykin and Network, 2004)—the focus of this survey paper—constitute one of the most popular ML models in the literature. ANNs are composed of artificial ‘neurons’ interconnected together in a structure that aims to mimic the neural processing (organization and learning) of biological neurons and its behavior. NNs seek to emulate the learning system of the human brain that itself consists of a large number of biological neurons interconnected in networks that govern all aspects of human behavior. ANNs are modeled on human brain and use a network of nodes called artificial neurons to form complex systems. The ways in which these nodes are placed and are connected lead to different kinds of NN models.

Broadly speaking, ML systems can be categorized into three kinds of learning systems: (i) supervised learning; (ii) unsupervised learning; and (iii) reinforcement learning. Using these learning systems, NNs can predict outputs for a set of given inputs.

Historically speaking, starting from the work on perceptrons, NNs were devised in a model that involved supervised learning. This means that the NNs learned using a labeled data set, for which correct outputs are given for inputs. NNs continued to evolve and mostly found applications in supervised learning settings. Techniques were introduced over time that allowed NNs to learn in unsupervised settings in which no labeled data was available. Most of the modern day advancements in NNs such as deep NNs make the use of unsupervised learning. This is better suited for real-world problems where labeled data set is not available. The applications of such deep NNs in wireless networks are yet to be seen. Such applications, however, may be dominant in the future. NNs have found applications in vast variety of engineering fields including image processing (Egmont-Petersen et al., 2002), control theory (Hunt et al., 1992), data analytics, digital communications (Ibnkahla, 2000) among various applications.

The strength of NNs lies in their ability to provide generalized solutions through an architecture that is able to learn to improve its performance. NNs comprise of a number of elements called nodes that are configured in layers and connected to other nodes by links with a certain weight. The behavior of NNs is determined through (i) the way nodes are connected together into a network, and (ii) the adaptive link weights between these nodes that are tuned by a learning algorithm. Simply speaking, NNs are essentially a network of weighted, additive values with “nonlinear transfer functions”.

In this paper, the applications of NNs in wireless networks are discussed. NNs have been used in various types of wireless networks ranging from wireless sensor networks (WSNs) to cognitive radio networks (CRNs) for a wide variety of applications such as signal classification, parameter optimization, and network management. NNs have diverse implementations and a particular kind of a NN may be appropriate for a particular wireless networking functionality. In the next section, the different types of NNs, along with their architectures and learning mechanisms, will be discussed followed by the wide applications of these NNs in wireless networks.

Contributions of this paper: This survey, to the best of our knowledge, is the first paper that has focused on a comprehensive survey of the applications of NNs in wireless networks in general. Previous papers have either focused on general wired networks (Frank et al., 1999, Yuhas and Ansari, 2012), or have focused only on specific wireless networking configuration such as WSNs (Kulkarni et al., 2011), MANETs (Forster, 2007), or CRNs (Pattanayak et al., 2012).

Organization of this paper: This survey is organized in the following way. To facilitate the reader, acronyms used in this paper are collected in Table 1 for convenient reference. In Section 2, necessary background about machine learning and NNs is provided. In Section 3, a broad survey of applications of NN technology in wireless domain is provided. In Section 4, the categorization of applications of NN is provided according to the type of the wireless network. Guidelines for using NNs in wireless networks are given in Section 5. The limitations and challenges associated with implementing NN in wireless networks are provided in Section 6. Finally, this paper is summarized with concluding remarks in Section 7.

Section snippets

Background: machine learning and NNs

In the following sections, the aforementioned different machine learning methods will be described along with the NN models that utilize these learning methodologies. An overview of the different kinds of NNs and the relationship between them can be seen in Fig. 1.

Neural network applications in wireless networks

Wireless networking encompasses a vast field and it is difficult to restrict its scope. Any paradigm that employs a wireless medium for communication can be considered a wireless network (Table 2).

Cognitive radios

Cognitive radios (CRs) (Mitola, 2000, Mitola, 2006) are intelligent communication systems that are able to sense their environment so that they could accordingly configure themselves to meet set goals. These set goals can vary from increasing throughput to energy efficiency to reliable communication (Table 3).

The motivation which has driven research in cognitive radios is the promise to allow communication devices to use idle frequency bands within the frequency spectrum allotted to other users

Guide to using NNs in wireless networks

In this section, we will provide some implementation guidelines to successfully use NNs (in supervised and unsupervised configurations) in wireless networks. In particular, we will discuss the relevance and suitability of different NN techniques for different tasks. We will discuss pitfalls and will highlight limitations and challenges in Section 6.

Pitfalls, limitations, and challenges

Though NNs are capable of modeling complex non-linear problems, they do have some drawbacks. One of the main criticisms of NNs is that they are a black box approach and do not provide any domain specific insights about the model/network has been developed (Karlaftis and Vlahogianni, 2011). In practice, one typically requires modeling techniques that will provide insight into how the features are correlated with the objective function. For instance, Classification and Regression Trees (C&RT) can

Conclusion

We have provided a survey of the applications of NNs in wireless networks. Due to the generality of the NN technique, it has been extensively applied in wireless networks. In addition to providing a self-contained introduction to common models and configurations of NNs, we have also provided a detailed survey of applications of NNs in wireless networks. We have categorized the applications of NN in wireless networks both according to the wireless networking configuration, and according to the

References (163)

  • R. Khan et al.

    Localization performance evaluation of extended Kalman filter in wireless sensors network

    Proc. Comput. Sci.

    (2014)
  • P. Lingras

    Comparison of neofuzzy and rough neural networks

    Inf. Sci.

    (1998)
  • Abu Alsheikh, M., Poh, P.K., Lin, S., Tan, H.-P., Niyato, D., 2014. Efficient data compression with error bound...
  • R. Agarwal et al.

    Survey of clustering algorithms for MANET

    Int. J. Comput. Sci. Eng.

    (2009)
  • Ahmad, U., Gavrilov, A., Nasir, U., Iqbal, M., Cho, S.J., Lee, S., 2006. In-building localization using neural...
  • Akinduko, A.A., Mirkes, E.M., 2012. Initialization of self-organizing maps:principal components versus random...
  • S.G. Akojwar et al.

    Improving life time of wireless sensor networks using neural network based classification techniques with cooperative routing

    Int. J. Commun.

    (2008)
  • D. Alahakoon et al.

    Dynamic self-organizing maps with controlled growth for knowledge discovery

    IEEE Trans. Neural Netw.

    (2000)
  • Altini, M., Brunelli, D., Farella, E., Benini, L., 2010. Bluetooth indoor localization with multiple neural networks....
  • S. Bahanfar et al.

    Neural networks for error detection and data aggregation in wireless sensor network

    Int. J. Comput. Sci. Issues

    (2011)
  • Bahl, P., Padmanabhan, V.N., 2000. RADAR: an in-building RF-based user location and tracking system. In: INFOCOM 2000....
  • Baldo, N., Zorzi, M., 2008. Learning and adaptation in cognitive radios using neural networks. In: 5th IEEE Consumer...
  • Barabas, M., Boanea, G., Dobrota, V., 2011. Multipath routing management using neural networks-based traffic...
  • Barbancho, J., León, C., Molina, J., Barbancho, A., 2006. SIR: A new wireless sensor network routing protocol based on...
  • Battiti, R., Le, N.T., Villani, A., 2002. Location-aware Computing: A Neural Network Model for Determining Location in...
  • Y. Bengio et al.

    Learning long-term dependencies with gradient descent is difficult

    IEEE Trans. Neural Netw.

    (1994)
  • Y. Bengio et al.

    Greedy layer-wise training of deep networks

    Adv. Neural Inf. Process. Syst.

    (2007)
  • D.M. Blei et al.

    Latent Dirichlet allocation

    J. Mach. Learn. Res.

    (2003)
  • D.M. Blei et al.

    Variational inference for Dirichlet process mixtures

    Bayesian Anal.

    (2006)
  • Boanea, G., Barabas, M., Rus, A.B., Dobrota, V., 2010. Design principles and practical implementation of a situation...
  • M.D. Buhmann

    Radial basis functions

    Acta Numer. 2000

    (2000)
  • Capka, J., Boutaba, R., 2004. Mobility prediction in wireless networks using neural networks. In: Management of...
  • G.A. Carpenter et al.

    The art of adaptive pattern recognition by a self-organizing neural network

    Computer

    (1988)
  • Ciresan, D., Meier, U., Schmidhuber, J., 2012. Multi-column deep neural networks for image classification. In: 2012...
  • Clausen, T., Jacquet, P., Adjih, C., Laouiti, A., Minet, P., Muhlethaler, P., Qayyum, A., Viennot, L., et al., 2003....
  • Cordina, M., Debono, C.J., 2008. Increasing wireless sensor network lifetime through the application of som neural...
  • Cordina, M., Debono, C.J., 2010. Applying an SOM Neural Network to Increase the Lifetime of Battery-operated Wireless...
  • Cortez, P., Rio, M., Rocha, M., Sousa, P., 2006. Internet traffic forecasting using neural networks. In: International...
  • Dehni, L., Krief, F., Bennani, Y., 2006. Power control and clustering in wireless sensor networks. In: Challenges in Ad...
  • Dharmadhikari, V., Gavade, J., 2010. An NN approach for MPEG video traffic prediction. In: 2010 2nd International...
  • Duda, R., Hart, P., Stork, D., 2001. Pattern Classification, vol. 2....
  • N. Enami et al.

    Energy based clustering self organizing map protocol for extending wireless sensor networks lifetime and coverage

    Can. J. Multimed. Wirel. Netw.

    (2010)
  • N. Enami et al.

    Neural network based energy efficiency in wireless sensor networksa survey

    Int. J. Comput. Sci. Eng. Surv.

    (2010)
  • Fazio, P., DeRango, F., Selvaggi, I., 2010. A novel passive bandwidth reservation algorithm based on neural networks...
  • Fehske, A., Gaeddert, J., Reed, J., 2005. A new approach to signal classification using spectral correlation and neural...
  • M. Figueiredo et al.

    Unsupervised learning of finite mixture models

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2002)
  • Forster, A., 2007. Machine learning techniques applied to wireless ad-hoc networks: Guide and survey. In: 3rd...
  • Frank, R., Davey, N., Hunt, S., 1999. Applications of neural networks to telecommunications systems. In: Proceedings of...
  • M. Gandetto et al.

    Use of time-frequency analysis and neural networks for mode identification in a wireless software-defined radio approach

    EURASIP J. Appl. Signal Process.

    (2004)
  • R. Gargi et al.

    Improving the performance of dynamic source routing protocol by optimization of neural networks

    Int. J. Comput. Sci. Issues

    (2012)
  • Cited by (68)

    • Deep Reinforcement Learning for QoS provisioning at the MAC layer: A Survey

      2021, Engineering Applications of Artificial Intelligence
      Citation Excerpt :

      However, the paper focuses only on old DL models as it was published three years ago. Ahad et al. introduced Neural Networks (NN) as a way to tackle challenges related to decision making or parameter optimization in wireless networks, e.g., the highly dynamic and unknown network environments (Ahad et al., 2016). One of the paper’s severe disadvantages is that it focuses on traditional Neural Network (NN) systems.

    View all citing articles on Scopus
    View full text