Page 258 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

250 CHAPTER 12 Computational Intelligence in the Time

should be scaled. Evolutionary computation algorithms can also be considered
at compile time to identify the optimal setting of controlling parameters over
execution time.
• Adaptive sampling. Good amount of energy saving can be achieved by
implementing an adaptive sampling strategy where the sampling frequency is
adapted according to the current needs of the application. By reducing
the sampling frequencydsometimes below the value granting signal
reconstructiondwecanalsoreducethebandwidthneededforcommunication[10].
Machine learning and fuzzy logic can be fruitfully considered here to control the
energy consumption of the system by adaptively acting on the sampling rates of
sensors by also taking into account predictions for both the residual and the
harvestable (in the future) energy.
• Keep the solution simple. This design strategy is always up-to-date in the sense
that, in general, complex solutions require high energy to carry out the due
computation which, most of time, is not needed. In fact, in presence of
uncertainty affecting the measurements and with the optimal application to be
executed on the CPS unknown, it does not make much sense to implement too
complex solutions. Machine learning and statistical methods should be inves-
tigated to assess the loss in performance associated with a given solution by also
taking into account existing uncertainty and available hardware resources.
• Consider incremental applications. Whenever performance accuracy is an issue,
we can tradeoff accuracy for energy, in the sense that if a higher accuracy level
is needed, then we tolerate the algorithm to be more complex and energy-eager.
Identification of the tradeoff between accuracy level and energy savings can be
carried out with optimization algorithms, for example, those based on
evolutionary algorithms.
• Duty cycling. The more you sleep (i.e., the device enters low and deep power
sleep modalities), the less energy you consume. By implementing duty cycling
at the processor, sensor, and memory levels, we can significantly control energy
consumption. Identification of timing to switch on and off different hardware
elements, as well as selection of the optimal sleep modality, represents an
application-dependent complex optimization problem whose optimal solution
can be found at compile time with both machine learning and evolutionary
computation algorithms.

4. LEARNING IN NONSTATIONARY ENVIRONMENTS

In designing cyber-physical and IoT applications we mostly assume that the process
generating the sensor datastream is either stationary (i.e., data or features extracted
from the signal are independent and identically distributed [i.i.d.] random variables)
or time invariant (the signal does not show an explicit dependency on time) [11].
However, such assumptions are hardly met in real applications and represent, in
the best case, a first-order approximation of the reality. In fact, sensors and actuators