Professional Documents
Culture Documents
Intelligence at The Edge of Complex Networks: The Case of Cognitive Transmission Power Control
Intelligence at The Edge of Complex Networks: The Case of Cognitive Transmission Power Control
Pasquale Pace and Giancarlo Fortino are with the University of Calabria — Rende; Yin Zhang is with Zhongnan University of Economics and Digital Object Identifier:
Law; Antonio Liotta is with the University of Derby. 10.1109/MWC.2019.1800354
Intelligent
Intelligent n - intermediary nodes
•Intermediary •Intermediary
Node Node
•Sensing and •Sensing and •Sensing and •Sensing and •Sensing and •Sensing and •Sensing and
Actuator Actuator Actuator Actuator Actuator Actuator P2P Actuator
Node (SAN) Node (SAN) Node (SAN) Node (SAN) Node (SAN) Node (SAN) Node (SAN)
DUMB DUMB DUMB
Intelligent Intelligent Intelligent Intelligent
problems in real time, even when disconnected gies play an important role when it comes to reli-
from central big data capability. Starting from this ability of communication, often setting stringent
challenging vision of decentralized intelligence, requirements on the energy efficiency and robust-
here we discuss the need to move the intelligence ness of the wireless systems, which makes a type
toward the edge of the network in which many of “on-the-fly” adjustment to the residual energy
different resource constrained devices are placed; level necessary in practical implementation.
then we explore the various ways in which the AI and particularly ML have considerably
latest AI developments, particularly machine evolved in the past decade, with a broad set of
learning (ML), may help address the complex methods and applications. Our intention in this
requirements of IoT communications, highlighting section is to explore which general software
the crucial role of predictive rather than deter- architectures for learning (i.e., traditional vs. dis-
ministic communications. First, we illustrate the tributed, shown in Fig. 1) may be employed in
software architectures that may be used for mov- cognitive wireless communications to support
ing intelligence toward the edge and transferring the well-known cognitive cycle in which the radio
meta-learned models. Next, we look at learning should be able to learn from its past actions.
mechanisms that can naturally be deployed in the
IoT node. Finally, we introduce an exemplary case Centralized Learning
study where ML is successfully used to find the Up until recently, the predominant method for
delicate balance between spectrum and energy employing ML has been the centralized one.
efficiency in the wireless sensor network. In the IoT context, this has been referred to as
This level of autonomy and self-adaptation cloud-assisted sensing, whereby all contextual
(through ML) within individual nodes is crucial in data needs to be transported to a central point
a context that is continuously changing in ways before it can be used to make prediction models
that conventional deterministic models fail to cap- and inference. The sensor nodes are “dumb” in
ture. In our prototyping study we wanted to ascer- the sense that their functional requirements are
tain the extent to which lightweight ML could limited to data collection and communication.
beneficially be used in cognitive communications. All intelligence resides in servers, which is where
Our findings, based on a range of diverse sensors sufficient computing resources are concentrated.
(some having as little as 20kB memory), show Centralized learning shares the pros and cons
great potential and encourage further studies in of centralization. It can rely on rich datasets and
this direction. However, in this work we only pres- powerful computations. However, communica-
ent the results related to a specific sensor plat- tion becomes the limiting factor, particularly con-
form due to space limits. sidering the projected scale of IoT systems and
their predominant use of wireless spectrum.
General Software Architectures for Learning
Since new attractive applications have caused the In-Node Learning
data volume in both fixed and mobile networks Recent development in AI has made it possible to
to skyrocket in recent years, it is foreseeable that miniaturize ML processes to the extent that online
the technologies available today will be unable to learning can be run on limited devices, with as
satisfy users’ demands for service quality in the little memory as a few dozen kilobytes [7]. Thanks
future. In this context, cognitive wireless technolo- to new breakthroughs at the intersection between
Convergence
0.1
0.01
0.1
0.1
0.8
0.8
600 k < 750
750 k < 900
Conclusion
Many powerful AI and ML techniques address-
Testing 0.01 0.01 0.8 k 900 ing knowledge engineering, planning, schedul-
ing, and learning in distributed environments
TABLE 2. Parameter setting for the phases of learn- have been developed in the last few years to
ing, convergence, and testing. face different networking and communication
issues taking advantage of the availability of
(MaxPow) is always used for the transmission. both processing and storing capabilities with-
As performance indices to validate the cogni- in the cloud. In the near future, attention on
tive approach within the presented commu- the design and application of such powerful
nication scenarios, we computed the average techniques will be moved from the core to the
values of PRR, latency, number of retransmis- edge of complex and heterogeneous networks
sions, number of CCA attempts and energy by adapting and re-designing their working
consumption of the whole network. We also features to face the limitation imposed by the
evaluated the convergence time of the CQL- presence of widely used and constrained IoT
TPC scheme in relation to the density of the devices.
network. This is computed as the time needed From this perspective, the presented work
by the agent to reach the Q-value that differs discusses the potential and, at the same
at most 1 percent from the final value at the time, the needs of moving more intelligence
end of each experiment. toward the edge of the networks by making
The obtained results are summarized in Fig. constrained devices evolve through the use of
4a, where it is possible to note that most of the lightweight ML and AI techniques, paving the
performance between CQL-TPC and MaxPow are way for the so-called actionable intelligence
similar in all the scenarios. In particular, the PRR at the edge. The implemented case study,
slightly drops when the network is composed of focused on the design of a cognitive transmis-
eight pairs of nodes; on the contrary, the latency, sion power control well suited for small real
retransmissions, and CCA attempts rise propor- sensors, proves the feasibility of the proposal
tionally to the number of nodes. in terms of energy efficiency, reliability, and
The main and significant difference that con- scalability, also ensuring standard network per-
firms the value of the cognitive approach is relat- formance levels.