Wireless networks are key enablers of ubiquitous communication. With the evolution of networking technologies and the need for these to inter-operate and dynamically adapt to user requirements, intelligent networks are the need of the hour. Use of machine learning techniques allows these networks to adapt to changing environments and enables them to make decisions while continuing to learn about their environment. In this paper, we survey the various problems of wireless networks that have been solved using machine-learning based prediction techniques and identify additional problems to which prediction can be applied. We also look at the gaps in the research done in this area till date.
In the current age of Information Technology revolution, quick availability of use of information to make speedy decisions is becoming a competitive advantage for many businesses. In such an environment, to have all decision-makers connected in, ubiquitous communications have become the need of the hour. Ubiquitous computing and communication combine mobility with context awareness, adaptability, scalability and localization to create an environment where devices are smarter and take actions by predicting user behavior. It finds its applications in a large variety of areas including energy conservation, manufacturing, healthcare, banking, education and telecommunications. In order to make ubiquitous computing a reality, certain devices are absolutely necessary. At the bottom of this ubiquitous computing stack, there are the sensors or smart phones with sensor functionality. These are responsible for collecting information from the surrounding environment and reporting it to a decision-making entity. In order to allow these sensors to communicate, the next layer is the wireless communication layer which can be provided by the 802.11 family of networks or any other communication technology. The final level in the stack includes the applications that collect, mine and analyze the data collected by the sensing devices for patterns in order to make decisions. Wireless networks, being a key enabler of the ubiquitous communication paradigm are gaining in importance. The quality of service provided by these networks is of utmost importance in determining the applications that they will be put to. Not only do these networks have to provide an enhanced quality of user experience (QoE), they must do so at optimized rates and with optimum resource usage. Wireless devices are extremely constrained for power and the network implementations have to account for this at all times.
With smart-phones being used for newer real-time applications every day, it becomes challenging for the wireless network elements to keep pace. It is thus the need of the hour for networking software to adapt to changing requirements and user trends without the need for manual intervention. One way of achieving this is to build intelligent network elements that record user behavior, characterize it, identify patterns in it and use the knowledge gained from these data to adapt various parameters. Foremost in the requirements of this type of intelligent software is the ability to predict different aspects of user behavior in order to determine any parameter changes or other actions sufficiently in advance, so that changing network conditions, due to mobility or other reasons, are seamless for the user. The use of artificial intelligence, specifically learning and prediction techniques make these adaptable systems a reality.
In this paper we look at the applications of prediction techniques to solve different aspects of the ubiquitous computing problem. The remaining part of the paper is organized as follows: in Section 2, we discuss machine learning techniques for prediction in more detail; Section 3 categorizes the literature in the area of prediction in wireless networks, based on the problems addressed using prediction; Section 4 and beyond look at each of these areas in further detail and compare the various methodologies used; finally, we look at the research gaps in this area of building intelligent wireless networks using prediction techniques.
Machine learning deals with algorithms that give computers the ability to learn, in much the same way as humans. This means that given a set of data, an algorithm infers information about the properties of the data, allowing it to make predictions about other data it may see in the future. The main focus of machine learning is the design of algorithms that recognize patterns and make decisions based on input data. Machine learning has found uses in areas like biotechnology, fraud detection, wireless networks, stock market analysis and national security.
Machine learning algorithms can be categorized into:
1) Supervised learning: these set of algorithms use training data to generate a function that maps the inputs to desired outputs (also called labels). For example, in a classification problem, the system looks at example data and uses it to arrive at a function mapping input data into classes. Artificial neural networks, radial basis function networks and decision trees are forms of supervised learning.
2) Unsupervised learning: these set of algorithms work without previously labeled data. The main purpose of these algorithms is to find the common patterns in previously unseen data. Clustering is the most popular form of unsupervised learning. Hidden Markov models and self-organizing maps are other forms of Unsupervised Learning.
3) Semi-supervised learning: as the name indicates, these algorithms combine labeled and unlabeled data to generate an appropriate mapping function or classifier.
Artificial neural networks are extremely popular in the field of prediction in wireless networks. The other techniques like decision tress and unsupervised learning are used much lesser. Experiments prove that using a combination of techniques instead of a single one provides the best results.
As wireless networks move towards being omnipresent to facilitate smart homes, offices and so on, completely eliminating any manual intervention in setting these up and operating them is essential. In order to adapt to the environment and inter-operate with other systems, making these systems intelligent will be essential for their success. One of the main problems in wireless networks is the unpredictable signal quality. The signal strength at any point in a wireless network is impacted by several factors—topology of the area, presence of buildings, interference from different networks and appliances operating at similar frequencies and so on. Since, the networks or the neighboring devices can keep changing, having a static algorithm to attack this problem will not work. Instead, having wireless network elements sense interference and respond to it appropriately will allow adaptation to changing environments. One of the most researched areas in ad-hoc and wireless sensor networks is in wireless link status prediction. Being able to predict when a link’s strength will drop below threshold levels and for how long, will allow applications to take corrective action in advance, thus ensuring minimal service disruption. Section 4 looks at the application of machine learning to this problem.
As the different types of networks and devices multiply, it is essential to allow inter-operability amongst these to give users maximum flexibility. Amongst other aspects, achieving inter-operability involves achieving seamless handovers between these networks. Different schemes exist to ensure a seamless handover between networks, but to complete the handover in time with minimal wastage of resources requires prediction of the time at which the mobile station will lose connectivity to its current network. This prediction of mobility and disconnection time is the focus of Section 5.
The most researched areas in ad-hoc and wireless sensor networks are those of routing and intrusion detection. The ad-hoc nature of these networks means that mobile nodes are responsible for routing packets while processing their own data, but users can enter and leave the network at any time. The main focus of research in this area is to have routing algorithms that can adapt to the changing topology as quickly as possible. Having some amount of prediction capabilities built into these algorithms allows them to select the most reliable and longer duration routes to forward data. Section 6 looks at the use of prediction in routing in further detail. The ad-hoc nature of these networks also makes them extremely vulnerable to security attacks. With the emergence of newer forms of security threats all the time, being able to use the data from previous attacks to predict the general characteristics of attacks and to detect newer ones will go a long way in securing these systems. Use of prediction in intrusion detection is the main focus of Section 7.
The stability and reliability of links in wireless networks is dependent on a number of factors such as the topology of the area, inter-base station or inter-mobile station distances, weather conditions and so on. As such, there is no single way of modeling wireless link behavior that will work in all cases, making it difficult to predict wireless link availability using mathematical models. At the same time, an estimation of link quality and link availability duration can drastically increase the performance of these networks, allowing the network to take proactive measures to handle impending disconnections. One such application of link disconnection prediction is discussed in [
Another scheme for link quality prediction and link estimation called 4C [
In order to predict wireless network connectivity, that is, the signal to noise ratio for a mobile station, [
In this section, we look at prediction to facilitate smooth handovers in further detail. Given the ubiquitous computing environment, together with the smart devices that can support multiple technologies and the application requirements to stay connected all the time, handovers across technologies are more widely supported and researched than in the previous generation of technologies. Handovers can be classified into:
1) Horizontal Handovers: these are handovers between same technology base stations.
2) Vertical Handovers: these are handovers between base stations belonging to different technologies, and as such are more challenging to handle than horizontal handovers. This is because of the variable handover times depending on the target network, together with the different procedures involved in handover.
Prediction to facilitate smooth handovers involves being able to predict the next location or point of attachment of mobile stations. Being able to predict the next location well in advance allows evaluation of candidate target networks to determine which one best meets the requirements, reservation of resources in the target network to avoid ping-pong of handovers and minimal loss of data since, the handover can be completed just as the mobile station loses connectivity to its current network. Prediction of the next location of a mobile station is also called mobility management. A lot of the literature in this area like [
Other schemes use information like network topology and delve deeper into user characteristics to be able to predict the user’s next location. One such technique is described in [
The BMP scheme is assumed to be implemented by a server which can be co-located with the authentication server. The server uses all of the user’s movement and location characteristics, together with the time of the day, to arrive at prediction lists consisting of next points of attachment for the MS. The MS then re-associates with the AP in the cells specified in the prediction list in order of their appearance in the list, if associations fail. During the next handoff, the first prediction in either list is used as the next cell prediction based on whether the duration of the MS in the current cell is medium or long. If the first prediction fails, the second is used and so on. A full scan is performed when all the predictions in the lists fail.
The authors compare the BMP technique to other techniques that are used to predict a user’s next location, such as determining next location based on signal strength, employing extra devices or an overlay network to detect APs and so on. BMP scores over these techniques because none of them can completely eliminate the need for scanning for APs, nor do they take the location topology or structure into account. The authors argue that the other prediction schemes that exist in the literature do not take the nuances of WLANs into account such as highly overlapped cells and MAC contention. In addition, the location-based schemes cannot capture mobility patterns that deviate from the norm.
Another category of literature that tries to solve the vertical handoff problem (handoff across technologies) is based on the IEEE 802.21 standard that defines a middle ware architecture to ease handoff across technologies, called Media Independent Handover Functions (MIHF). The architecture defines the Event Service which provides information about change in local and remote link layer conditions in the form of events and triggers. These triggers include: 1) Link Up (LU), 2) Link Down (LD), 3) Link Going Down (LGD), 4) Link Going Up (LGU). The LGD trigger leads to the network triggering a handover for the corresponding mobile station. Receiving this trigger too late means that the link will be lost before the handover is complete and hence, there will be data loss. Receiving this trigger too early means that there will be a wastage of network resources, since the link was still working in the source network when the handover occurred. Hence, the right timing of these triggers is essential for an efficient handover. Thus, a large body of literature in this area attempts to find different mechanisms to predict the timing of these triggers. We discuss some of it in this section.
The algorithm presented in [
As the signal strength approaches this level, the prediction of the source network signal strength starts. The prediction process ends when the signal is lost. At this point, the MS should be able to connect to the target network for a seamless HO. Hence, the time taken to initiate and execute the HO is taken into account to calculate the time at which the prediction of the target network signal strength starts. This prediction continues until the predicted signal strength crosses the level at which the MS can safely connect to the target network without losing any data. The authors compare this scheme to the scheme used in [
A cross-layer predictive handover architecture based on the 802.21 standard is proposed in [
Ad-hoc networks and wireless sensor networks are growing in popularity because of the limited infrastructure needed to make these networks a reality and the self-organizing nature of these networks. The self-organizing nature of these networks is an area of active research due to the need for the networks to mimic the operator’s intelligence in configuring and maintaining themselves. One such prominent research area is that of packet routing in these networks. Since, these networks work based on different mobile stations serving as intermediate hops, changes in topology and thus routing paths are very frequent. Routing algorithms that adapt to these changing topologies while consuming minimal energy are an important requirement in these networks. This section looks at how prediction techniques can be used to overcome some of the routing problems in ad-hoc and sensor networks.
In [
Cognitive radio is a technology devised to overcome the spectrum shortage problem, by allowing unlicensed users (also called Cognitive Users—CU) to use the unused parts of the spectrum originally allotted to Primary Users (PU). Towards this end, individual nodes sense their environment and adjust their transmission parameters to minimize interference with primary users. While this allows extremely efficient spectrum utilization, it makes routing a challenge in these networks, where the chances of interference with primary users are much higher and thus the links are unreliable and available for shorter durations. The application of prediction-based algorithms to the problem of topology control and reliable routing in cognitive radio networks is discussed in [
1) Link availability time, predicted using the scheme proposed in [
2) Period of time spent in re-routing,
3) Link data rate.
This reliability metric is used to determine the weight of a link and the weight of a path. The algorithm then proceeds to construct the topology, by identifying all neighbors, estimating the path weights from initial node to unvisited nodes and then constructing the complete topology using the paths with maximum weight. This ensures that re-routings are minimized because using the path weight equation, links with high data rate and low availability time and links with low data rate and high availability time are avoided as far as possible. This allows the routing algorithms to indirectly take into account the mobility of CUs as well as the interference from PUs in routing decisions. Through simulations, the authors prove that the resulting routes are more reliable and lead to lesser re-routings.
The algorithm proposed in [
The emergence of ad-hoc and wireless sensor networks has brought in several advantages like efficient utilization of the spectrum, reduction in Capital Expenditure (CAPEX) due to the absence of a fixed infrastructure, reduction in Operating Expenditure (OPEX) because of their self-configuring nature and allowing much better monitoring of military areas and making wireless body area networks and ad-hoc vehicular communication a reality. However, greater flexibility also makes these networks more vulnerable to security threats and attacks. Hence, different authentication mechanisms, attack detection and attack prevention mechanisms have been studied extensively. However, increase in the computing power allows large equations and passwords to be broken in a matter of seconds and hence, security algorithms need to keep evolving to keep finding and fixing newer and newer vulnerabilities. In all cases of network security, detection of an attack or a security threat is the biggest challenge. In this section we look at the use of machine learning techniques to detect intrusion and denial-of-service attacks in ad-hoc and wireless sensor networks.
The algorithm presented in [
1) The supervised learning layer resides in the individual sensors. A decision tree is used as a classifier to perform this supervised learning. The decision tree is contained within the detection agent. This agent uses a set of rules to drive the classification process of the tree. The results of the classification are further used to update the rules. If an attack that is unknown to the sensors occurs, it is sent to the sink node.
2) Unsupervised learning performed at the sink. A decision tree is used to perform clustering at the base station. If an attack unknown to the sink occurs, it is propagated to the base station.
3) Reinforcement learning performed at the base station: reinforcement learning is used to predict intrusion in advance. The authors use a convergent temporal-difference learning scheme [
The database agent logs all events and attacks and provides an interface for querying by the detection and prediction agents. The communication agent facilitates communication between the sensors, sinks and the base station.
The authors evaluate the time overhead, memory consumption and communication overhead of this scheme in addition to its prediction accuracy. The algorithm is found to have the lowest time overhead as compared to the schemes proposed in [
A scheme to detect malicious nodes based on energy prediction is proposed in [
1) Selective forwarding attack: in this case, the energy dissipation is lower than the predicted value.
2) Hello flood attack: substantially higher energy dissipation than predicted.
3) Sybil attack: difference in energy consumption is larger than a preset threshold.
4) Wormhole attack: double the predicted energy is consumed.
5) Sinkhole attack: difference in energy dissipation increases gradually.
Simulation results show that the scheme is more efficient than existing ones in that it has high prediction accuracy and does not require any monitoring at individual sensor nodes. As a result, it can detect attacks with the least energy consumption which is ideal for limited-resource networks like wireless sensor networks.
A large portion of the research in wireless networks relies on knowing the status of the link as reliably as possible. The wireless link prediction schemes predict the time at which the signal quality will degrade. However, these are looked at mostly from a handover perspective. Several times, the signal quality degradation is a transient condition because of environmental factors and do not eventually lead to a handover. As such, using historical data to predict disconnection duration would be instrumental in a lot of applications like video streaming, browsing sessions based on TCP and so on. Being able to predict the disconnection duration in non-handover scenarios is thus a gap which needs to be researched in more detail.
Most of the research in the area of imparting intelligence to network elements uses artificial neural networks to perform prediction. With machine learning itself being an active area of research, a lot of newer models have been formulated and several experiments run to prove that these work better than artificial neural networks. Techniques like Random Forests and Deep Learning have been proven to achieve high prediction accuracies. Applying these techniques to the current wireless network problems and comparing their accuracy against that of artificial neural networks will help to establish the feasibility of these techniques and set the stage for their usage in wireless network research.
The other area that needs to be looked at in further detail is the efficient implementations and resource consumption of machine learning techniques in real-time devices. While there is a large body of literature that uses machine learning techniques to solve problems in wireless networks, only a small portion of it actually looks at how efficient and resource-usage friendly each of the techniques is. Given the computation intensive nature of some of the algorithms, it is perfectly possible that they lead to extremely accurate predictions but cannot be employed in any of the devices because of the resource consumption involved. Evaluating the various techniques based on their resource consumption in different systems is thus a topic that needs further work. In addition, research in the area of building efficient implementations of machine learning techniques for wireless networks, which take into account the limited memory, computing power and battery life in these networks is imperative.
Finally, all machine learning models are heavily dependent on the availability of real datasets. Today, there are few datasets available [
An ever-increasing customer base and the need for ubiquitous computing pose new challenges to network operators. The network elements must be able to continuously evolve with the user demands. This is only possible if they are designed to adapt to changing network conditions. Building adaptability into a system involves providing it with some level of intelligence that enables it to take decisions in different situations. In this paper, we looked at the application of prediction techniques to different wireless network problems like handover latency reduction, routing, link duration prediction and so on. In most cases, these techniques provided a significant improvement over their static counterparts. The large body of research in this area also indicates that the industry is slowly but surely realizing that systems have to be more and more adaptable in order to be able to handle the data explosion. As more machine learning techniques evolve, researchers are beginning to look at more un-or- thodox techniques that give higher prediction accuracy and are less performance-intensive. A considerable amount of effort is also being put into adapting existing techniques for use in real-time systems.