IEEE - Aerospace and Electronic Systems - September 2021 - 47

Martone et al.
uncertainty [33], realizing a decision rule (or policy) that
maximizes the expected value ofa user-defined reward based
on an application-specific combination of SINR and bandwidth.
Recent research examines Deep Q-Learning, which
can handle a larger problem space than the MDP algorithm,
including additional memory states [34]. An adversarial bandit
model has also been examined which maintains favorable
performance even in the presence of intelligent interference
sources (i.e., intelligent jammers) [35]. In [36], the SPA and
SLA CR techniques were compared and found to yield similar
performance in many stochastic scenarios.
In contrast to the avoidance approaches that seek conFigure
2.
Example spectrum with low and high power RFI. FSS refines the
number of frequency bins handled by a decision process.
ofwhich six are low power and five are high power. The second
function, " Merge, " then merges high-power groups into
spectral clusters to be avoided depending on the waveform
structure to be employed (e.g., determining the single largest
contiguous bandwidth or via spectral notching).
The output ofFSS is the metadata comprised of center
frequencies, bandwidths, and power levels of the merged
clusters. This metadata then informs the operation of the
particular CR technique to be deployed. The techniques
that have been explored thus far include sense-react-avoid
(SRA), sense-predict-avoid (SPA), sense-learn-avoid
(SLA), sense-react-notch (SRN), and sense-predict-notch
(SPN), with the spectral parameters determined by each
used to synthesize a waveform for DSA.
The SRA technique uses the FSS metadata to quickly
react to RFI by selecting the center frequency and bandwidth
ofthe particular low-power group having the largest contiguous
bandwidth [30], [31]. The SPA approach likewise seeks
the largest bandwidth, but does so in an anticipatory manner
by employing a stochastic model that represents the RFI as
an alternating renewal process [32], with the RFI statistics
learned online in real time. Similarly, SLA is a reinforcement
learning approach that treats the CR's waveform selection as
aMarkov decision process (MDP) for decision making under
Table 1.
Summary of CR Techniques for DSA
Technique TCR
SRA
SPA
SLA
SRN
SPN
SEPTEMBER 2021
tiguous bandwidth, the SRN approach uses the FSS metadata
to inform subsequent spectral shaping optimization
that realizes instantiations of random FM (RFM) waveforms
that are continuous, have a constant time-domain
envelope, and possess an enforced spectral mask containing
notches [37], [38]. The main advantage of this
approach is that the full bandwidth can be utilized, sans
notched regions, though a higher computational cost is
incurred for waveform generation. It has also recently
been shown that SRN can likewise be combined with prediction
[39] to facilitate the CR technique of SPN, which
addresses real-time computational demand as well reducing
the achievable latency in responding to RFI changes.
A summary of these CR strategies, all of which can
adapt on a pulse-to-pulse basis, is provided in Table 1.
The RT ofeach method is denoted using TCR, which measures
the processing time (or latency) that is quantified
according to the implementation strategy and hardware
discussed in the section " Response Time. " The Performance
Metric denotes the means used to evaluate each
CR technique, where the prime metrics ofSINR and bandwidth
are self-explanatory, while collisions (CO) and
missed opportunities (MO) respectively indicate unintentional
mutual interference and available spectrum not utilized
by the radar. The Clutter Modulation performance
metric evaluates the phenomenology (section " PAC Regulation " )
that arises when subcoherent processing interval
(CPI) waveform adaptations occur, which causes pulse
compression sidelobe and mainlobe modulation across
Performance Metric
Response Strategy Ref
164 ms SINR / Bandwidth / Clutter Modulation Reactive
410 ms CO / MO
410 ms SINR / Bandwidth
Predictive
Predictive
451 ms CO/ SINR / Bandwidth/ClutterModulation Reactive
451 ms SINR / Sidelobes
IEEE A&E SYSTEMS MAGAZINE
Predictive
[30], [31], [40], [41]
[32], [36]
[33]-[35]
[37], [38], [42]
[39]
47
IEEE - Aerospace and Electronic Systems - September 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - September 2021
