IEEE Circuits and Systems Magazine - Q3 2019 - 38

are examples of stochastic patterns. A deterministic
example is radar patterns that are formed by the periodic rotation of the radar antenna, allowing secondary
use of the spectrum by cellular users [14]. Classification
techniques for traffic patterns include the use of autocorrelation data, maximum likelihood classifiers, and
average likelihood functions [13].
Temporal availability information is often needed
to find the best channels. Both idle and busy time information is important to know how to decide which
channels to use and where to stay. Moving averages
of idle times can be used in stochastic traffic scenarios. Also, delay in finding a suitable channel should be
considered in spectrum sensing and allocation algorithms. For example, a limited set of channels decreases sensing time and may provide more time for actual
data transmission.
Decision-Making on Whether to Stay in the Channel
or to Make a Frequency Switch
Sometimes it might be a good decision to stay in the
current channel instead of switching to a new one when
the incumbent user appears [15]. A frequency switching
incurs delay and requires overhead in channel setup.
Thus, a secondary spectrum user may choose to stay
silent in the evicted band for future reuse if the primary user transmission duration is predicted to be
relatively short. A secondary system should have a
learning strategy to be able to minimize disruption time
of the own data transmission on-the-fly and be able to
decide whether to stay or move.
Evacuation times with current cellular base stations,
e.g., in LSA and CBRS/SAS scenarios, strongly affect the
ability to use the cellular networks for secondary users
of the spectrum [7], [8]. A secondary system should be
able to shut down transmission in the current channel
rapidly without causing interference to incumbent users. Ability to predict the incumbents' appearance relaxes requirements but does not eliminate them.
Predictive Network Slicing
Applications have their own quality of service (QoS) requirements that need to be supported by the network.
Mission-critical data often have very strict latency requirements that cannot be supported with conventional
network approaches [16], [17]. Thus, an interesting idea
is to learn the characteristics of current network traffic
and use proactive dynamic slicing techniques to dedicate temporarily part of the network resources to these
applications. This approach could be considered as
predictive QoS slicing for 5G networks where a network
slice is a complete logical network providing telecommunication services and network capabilities.
38

IEEE CIRCUITS AND SYSTEMS MAGAZINE

Joint Spectrum and Mobility Prediction
The number of handovers will increase in dense multiRAT 5G environments and mobility management will be
challenging. It is possible to use historical information
of previous locations to predict where the wireless device is going and make proactively seamless handover
decisions. In addition, predictive resource allocations
along the device's route can be made based on the requirements of the used applications. In typical mobile
networks, nodes exhibit some degree of regularity in
mobility patterns. For example, a car traveling on a road
is likely to follow the path of the road and a student
walking inside a university is likely to continue along the
same corridor. Broadly speaking, it is possible to learn
how to predict the future mobility behaviours relying
only on the mobility history (e.g., speed and direction)
and then use that information in predictive local-aware
spectrum allocations. This is even more important currently due to the decreasing size of 5G small cells.
Prediction of spectrum use can be combined with
location-awareness and mobility prediction. The ability to predict the location and mobility of devices while
taking recent historical information about the usage of
radio resources enables efficient allocations [18]. For
example, spectrum resources for critical information
of health devices can be allocated proactively based on
local channel fluctuations and on the predicted path of a
moving device so that low-latency communications can
be ensured at any location inside a hospital. Another
example in outdoor scenarios is predictive support for
autonomous vehicles and ships that need multiple sensors (laser range finders, radars, optical cameras, infrared, etc.) to obtain knowledge about their environment
when driving on the roads or operating at sea [19].
In addition to own location information and route,
it is essential to know the locations and movement of
others in the area in advance to avoid collisions and
operate safely. This requires reliable communication
methods and location-aware predictive resource management for autonomous driving and sailing. A third
example is a spectrum sharing scenario between a mobile satellite system and a cellular system operating in
the same frequency. The movement of the satellite can
be used to predict a possible interference situation and
start antenna tilting, power adaptation, or frequency
change process in order to enable an interference-free
operation for both systems.
Deep Learning
Deep learning [20] may provide methods and tools to realize highly dynamic operations and rapid autonomous
learning of the operational environment. According to [4]
the main learning problems in CR systems are decision
THIRD QUARTER 2019

IEEE Circuits and Systems Magazine - Q3 2019

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q3 2019