IEEE Circuits and Systems Magazine - Q3 2019 - 37

using spectrum sharing is partially dependent on the
observation time and memory length. However, the
speed of operations is usually slower the more intelligent the system is, as depicted in Fig. 2. The lowest
hierarchy level can be simple and operate in real time,
but the upper hierarchy levels are more complicated
and operate at a lower speed [9]. Intelligence and history information can be used in numerous ways. We will
classify the application areas into three groups, i.e. a)
spectrum sensing and measurements, b) radio resource
management and channel selection and c) joint spectrum and mobility prediction. We discuss the characteristics of each in the following.

the seminal paper of [1] was published. It is quite good
metric in long term studies, but pure occupancy-based
selection does not guarantee selection of good channels in all scenarios. Multi-criteria decision-making and
trade-offs are needed in many practical scenarios since
not a single metric is the best in all scenarios. An example trade-off is between the sensing duration to find
the best opportunities and the sensing overhead caused
by frequent sensing. Another multi-criteria example is a
combination of location and spectrum to support predictive local-aware spectrum allocations. Multi-objective optimization methods provide means to find Pareto
optimal solutions [12].

Sensing and Measurements
Spectrum occupancy studies aim at quantifying the proportion of time when a certain frequency channel or frequency band is occupied in a given area, describing their
utilization rate [3]. When the data is gathered over longer
time periods such as days or weeks, one can make long
term predictions to support spectrum management and
regulatory decisions. Long term predictions show how
different bands are used at different times, which helps
in finding both under- and overused bands and make rational decisions for future use.
Big data techniques such as machine learning and cloud
computing are important especially in long term spectrum analysis. For example, in a continuous multi-year
study described in [3] the amount of stored data daily is
3 GB. This adds up to a very large storage requirements
over weeks and months of measurements. Big data analytics is needed to find relevant temporal patterns. One
can use either local or cooperative data in prediction.
Collaborative prediction is shown to achieve better accuracy in most traffic conditions but may add more complexity to data analysis. A predictive approach, which is
able to exploit the predictable information from big data
explicitly, is used in reducing energy consumption of cellular networks in [10].

Radio Resource Management and
Channel Selection
In general, channel selection approaches in DSA systems
can be classified as 1) Underlay transmission in a channel where the primary user is operating. Transmission
power is limited by interference constraints. 2) Reactive
selection of a new available channel when needed, e.g.,
when an incumbent appears in the current channel. 3)
Predictive channel selection where historical information is used to assist the selection. This section shortly
discusses some important prediction concepts and the
interested reader is encouraged to look at the references
[2]-[6] to obtain more details about techniques such as
Bayesian approaches, Hidden Markov Models, or timeseries analysis that can be applied both for short term
and long term usage prediction.

Assistance for Sensing and Optimization
of Sensing Periods
Long term predictions can be used to focus sensing on
the most promising channels in spectrum sharing scenarios [11]. When the sensing results of different channels are stored in the sensing history database, the decision engine can check what channels are most likely to
be free at the requested time (e.g., at 5 pm on Thursday).
Then, the sensing is focused on these channels to reduce total sensing time.
An essential part of the predictive DSA operation is
the use of good metrics. Increase of spectrum occupancy
has been one of the main targets of DSA research since
THIRD QUARTER 2019

Selection of the Best Channels for
Data Transmission
The main idea in predictive channel allocation is to be
able to select not only the instantaneously best channels for transmission but rather be able to learn and
select channels that will remain good in the future.
Predictive channel selection can be used to focus the
data transmission on channels that are estimated to
be available for the longest time periods. To be generally usable and accurate in different frequency bands
and scenarios, the predictive method should be able
to classify the traffic patterns of different channels
[13] and apply specific methods for these channels
[2]. A good metric to be included in the channel selection process is the channel switching time that strongly affects to the decision whether to stay or change
the channel.
Different prediction techniques apply to stochastic
and deterministic traffic patterns. In the former case,
the traffic can only be described by statistical terms
whereas in the latter exact predictions are possible. Mobile networks and typical voice and data applications
IEEE CIRCUITS AND SYSTEMS MAGAZINE

IEEE Circuits and Systems Magazine - Q3 2019

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