Scribd
Upload
29 views2 pages

Summary

The document discusses using regression models and deep learning techniques to predict packet delivery ratio (PDR) and energy consumption (EC) based on communication parameters in IEEE 802.15.4 networks. A public dataset containing measurements of over 48,000 parameter configurations is used to evaluate prediction performance. Accurate predictions of PDR and EC based on parameters can help configure networks to meet performance goals while balancing energy usage.

Uploaded by

saveii saveee
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
29 views2 pages

Summary

The document discusses using regression models and deep learning techniques to predict packet delivery ratio (PDR) and energy consumption (EC) based on communication parameters in IEEE 802.15.4 networks. A public dataset containing measurements of over 48,000 parameter configurations is used to evaluate prediction performance. Accurate predictions of PDR and EC based on parameters can help configure networks to meet performance goals while balancing energy usage.

Uploaded by

saveii saveee
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Reliability is among the key performance requirements for many loss-sensitive IoT and WSN

applications, while Energy Consumption remains a primary concern in WSN design. With the ever-
increasing availability of performance data, data-driven techniques are becoming popular in
such settings. It is observed that a number of communication parameters like transmission
power, packet size, etc., influence metrics like PDR and EC in diverse ways. In this work, different
regression models including linear, gradient boosting, random forest, and deep learning are used
for the purpose of predicting both PDR and EC based on such communication parameters.

To evaluate the performance, a public dataset of the IEEE 802.15.4 network, containing
measurements against more than 48,000 combinations of parameter configurations, is used.
These prediction results can help configure communication parameters taking into account the
performance goals.

Introduction

WSN-based IoT for smart buildings include energy and water management, fire and smoke
detection, appliance and light control, security and surveillance, and entertainment . IEEE
802.15.4 is one of the most popular communication standards used in WSNs. According to a
recent survey, WSN deployments for which communication standards are known, more than 50%
are based on IEEE standards . Recently, there has been a tremendous research effort to improve
all possible performance dimensions of WSNs and IoT .

Optimizing multiple metrics is frequently modeled as an optimization problem that is often NP-
hard . Optimizing one metric also tends to conflict with the other metrics, and a trade-off is
needed. Mathematical programming based serialization methods and nature-inspired meta-
heuristic algorithms are popular choices in this context . In addition, adaptivity is considered a
mandatory trait for all modern communication systems.

Hence, there is a growing trend of using data-driven approaches and Machine Learning to meet
the performance challenges faced by WSNs and IoT.

The focus in this paper is predicting reliability in the form of Packet


Delivery Ratio and

Energy Consumption in IEEE 802.14.5 based networks using Deep Learning . It is argued that
various communication parameters e. Findings from this study can help extend the research
effort encompassing a wider range of parameters including even protocols at different layers. In
this paper, an adaptive system to predict PDR and EC based on different stack parameter
configurations is proposed.

The system used performance data and applied deep learning to make data-driven predictions
for both PDR and EC. In particular, the suggested design achieves an adequate prediction
accuracy without having to solve any intractable optimization problem. IEEE 802.15.4 network.
This data can be directly used for choosing values of communication parameters, meeting the
constraint for metrics under consideration.

Desirably accurate estimation of performance metrics without having to deal with the
optimization problems which are intrinsically NP-hard. A flexible and evolving system that can
adapt to the circumstantial and even design changes that may occur over time. This
adaptiveness can facilitate a sustainable system in contrast to the most client-side approaches
where learning is based on the missing value prediction in the output matrix rather than the
input communication parameters. Prediction results are presented in Section 4.
With the benefit of live decision-making comes the drawback of the limited capacity of IoT
devices, thus compromising on the quality of learning as well as overloading an already
constrained device. , whereas uses a deep neural model. All these works are evaluated using
standard regression metrics and most of these use an established dataset released by Zheng
containing the RT and THP of 339 users and 5825 services.

Server-Side Predictions

Server-side approaches put the load of computation on a remote server with end-devices
primarily utilizing the recommendations from the server. Having adequate processing capacity,
sophisticated learning algorithms can be applied to large-scale performance data. Although the
computational burden is eased, a certain amount of periodic communication overhead has to be
afforded in such settings. Some of the main server-side prediction approaches are presented in
Refs.

Ayhan used neural networks to predict TP level in relation to network Lifetime and inter-node
DT. In order to predict one of these metrics, two others were taken as features in this work . The
client-side approaches are limited to missing value predictions from the matrix of values for the
metric under consideration. This limits the quality and applicability of prediction models to the
diverse application scenarios and heterogeneous communication settings present in the IoT.

It is argued that a model capturing the variations in different communication, circumstantial and
application specific variables, affecting the QoS, can better prepare a model for sustainable
learning and potentially yielding more accurate predictions. The server-side approaches, on the
other hand, do not take into account a wider set of parameters and often use one predicted
metric as a feature to predict another metric , or the values directly translating into the metric of
interest are used as input features . The realization of such a system is far from being practical.

Overview

A server collects this data and runs deep learning algorithms to identify the relationships
between performance metrics of interest and communication parameters , as highlighted in
Figure 1. This learning process keeps adapting as different aspects of the network evolve , and
new data becomes available. Against each performance metric constraint, a set of
recommended values for the communication parameters that meet the required goals for the
metric are sent to the central controller within the consumer premises. This table contains a set
of values for each metric and recommended values for a list of relevant parameters that may
help meet the constraint for the metric under consideration.

The controller at the user site uses this information to select the suitable values for the
parameters considering performance constraints.

A list of abbreviations and symbols is provided in Tables 2 and 3 summarizes these parameters
along with their explanation and the range of values used in the experiments.

Powered by TCPDF (www.tcpdf.org)

You might also like