Page 261 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

4. Learning in Nonstationary Environments 253

interested in knowing that a change in stationarity occurred (the change in
stationarity might be associated with a faulty sensor, for instance, as investigated
in the next section). If the computational complexity of the update phase is not
negligible, also in relationship with the dynamics of the change, we might prefer
active solutions. However, active approaches suffer from false alarms (false
positives) introduced by the change-detection triggering mechanism; thus inducing
the application to update even though not strictly needed. Fortunately, in CPS
applications false positives do not negatively affect performance but only introduce
an extra, not requested, computation.
One should expect that if the physical environment is changing with a low
frequency then an active approach might be more appropriate than a passive one.
However, again, one should balance the application update phase with its
computational complexity and the level of time variance exposed by the
environment the system interacts with. This problem becomes even more relevant
and up-to-date in CPSs, IoT, and Smart-X technologies, producing high-
dimensional datastreams where the computational load is expected to be high. Since
active and passive solutions represent extreme strategies, current solutions are
investigating hybrid approaches that aim at taking major advantages from both these
solutions.

4.3 RESEARCH CHALLENGES
Several research challenges should be addressed in order to support a quick and
effective design of cyber-physical and IoT technologies
• Design methodologies. Neither investigations nor methodologies are available to
shed light on the relationships among the effectiveness of active/passive
approaches with respect to the speed/nature of the change and the computational
complexity of involved methods. Such investigations are fundamental to permit
embedded applications to detect possible changes in the environment and react
accordingly to keep the quality of service at the appropriate level.
• Design of distributed decision-making applications. There are no computation-
ally light change detection mechanisms for distributed embedded systems able
to control the false positive rates. The challenge here is to provide
distributeddpossibly autonomousddecision-making strategies, reasonably
based on machine learning methods.
• Approximate computing. What is largely needed are strategies permitting the
harmonization of the learning in nonstationarity environment functionality for
distributed embedded systems, possibly within an approximate computing
framework. In this case, the approximation level introduced by the hardware, as
well as that introduced by the adoption of incremental software, should be
traded off with accuracy performance as coming from active or passive learning
strategies. It is expected that the optimization problem identifying the most
suitable level of approximation over time can be carried out with evolutionary
computation algorithms.