International Journal of Trend in Scientific Research and Development (IJTSRD)

Volume 4 Issue 3, April 2020 Available Online: www.ijtsrd.com e-ISSN:

ISSN: 2456 – 6470

A Generalized
lized Regression Neural Network
Approach to
o Wireless Signal Strength Prediction
Finangwai D. Jacob,
Jacob Deme C. Abraham, Gurumdimma Y. Nentawe
Department of
o Computer Science, University off Jos, Jos, Nigeria

ABSTRACT paper Finangwai D.

How to cite this paper:
This study presents a Generalized Regression Neural network (GRNN) Jacob | Deme C. Abraham | Gurumdimma
Gurumdi
based approach to wireless communication network field strength Y. Nentawe "A Generalized Regression
prediction. As case study, the rural area between the cities of Bauchi and Neural Network Approach to Wireless
Gombe, Nigeria, was considered. The GRNN based predictor
pr was created, Signal Strength Prediction" Published in
validated and tested with field strength data recorded from multiple Base International
Transceiver Stations at a frequency of 1800MHz. Results indicate that the Journal of Trend in
GRNN based model with Root Mean Squared Error (RMSE) value of 5.8dBm Scientific Research
offers significant improvements over the empirical Okumura and COST 231 and Development
Hata models. While the Okumura model overestimates the field strength, (ijtsrd), ISSN: 2456-
2456
the COST 231 Hata significantly underestimates it. 6470, Volume-4
Volume |
Issue-3,
3, April 2020, IJTSRD30501
KEYWORDS: Field Strength; Generalized Regression Neural Network; pp.467-471, 1, URL:
Okumura Model, COST 231 Hata Model www.ijtsrd.com/papers/ijtsrd30501.pdf

Copyright © 2020
20 by author(s) and
International Journal of Trend in
Scientific Research and Development
Journal. This is an Open Access article
distributed under
the terms of the
Creative Commons
Attribution License (CC BY 4.0)
(http://creativecommons.
(http://creativecommons.org/licenses/
by/4.0)
INTRODUCTION
In recent times,, soft computing based methods (also The reduction in intensity sity of the signal from its
termed computational intelligence) are used to solve a transmitting to receiving station is known as attenuation
variety of problems. This is as a result of their outstanding [7].. The attenuation of signal in wireless systems prompts
prompt
abilities
ies to mimic the processing power of the human radio engineers
ineers to plan adequately for modeling of radio
brain [1]. These techniques encompass the convolutional propagation for the purpose of signal prediction in view of
neural network (CNN), generalized regression neural establishing good network coverage.
coverage Radio propagation
network (GRNN), multi-layer
layer perceptron neural network models are applicable within a specific terrain.
(MLP-NN), adaptive neuro- fuzzy Inference In system
(ANFIS), radial basis function neural network (RBF-NN)
(RBF There are existing models that have been used widely for
etc. The aforementioned techniques are useful in areas like modeling radio propagation for accurate signal strength
signal prediction, image processing, pattern recognition, prediction in telecommunication systems. Some of these
voice recognition and so on. Soft computing techniques are models were created in Japan and Europe based on
capable of handling complex function approximation empirical data. Some of the broadly used models include
problems with greater accuracy than regression methods. deterministic and empirical models.
models [8]
In recent times, computational intelligencegence techniques
have been applied to the field of telecommunication for Deterministic models are based on ray tracing, which is
radio propagation modeling. suitable for predicting signal
nal strength in wireless systems
within short distances [9]. The accuracy of the model is i
Signal strength determination is one of the paramount due to its detailed requirements of information about the
requirements considered when planning wireless environment [10]. However,, the model is time consuming
telecommunication systems. This is because as signalssi in terms of computational effort.
effort On the other hand,
propagate through space, they tend to reduce in strength empirical models are preferable for radio propagation
due to variable factors such as diffraction, refraction, modeling as a result of their simplicity.
reflection, scattering, free space losses etc.
etc [2]. This is
termed as multiple path propagation [3].[3 The losses in Empirical models are mathematical
mathematic equations that are
signal strength of wireless systems can also be attributed based on in-depth
depth field measurements
measurement and observations
to the distance of receiving station from the transmitting [3], [11]. These models require less computational effort,
station, radiated power of the transmitter,
er, the height of implying greater efficiency in computation but not as
the antenna, trees, nature of the terrain etc. [4], [5], [6]. accurate when compared with deterministic models. Some

@ IJTSRD | Unique Paper ID – IJTSRD30501

30501 | Volume – 4 | Issue – 3 | March-April
April 2020 Page 467
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
of the broadly used empirical models include the upon sample values Xi and Yi of the random variables x
Okumura, Hata-Okumura and Cost 321 Hata etc. and y, where n is the number of sample observations and
is the dimension of the vector variable x.
x
This study is aimed at using an artificial neural network
called the GRNN to predict signal strength across the rural ,  
>
  ##<</ = <</  AA/ 
. ∑$% exp " ? . exp @ B (2)
area between the cities of Bauchi and Gombe, Nigeria.
Nigeria The
GRNN based model is created and nd compared with some of /  /  σ σ
the widely used empirical models such as the Hata-
Okumurah and the COST 321 Hata,, which are suitable for The scalar function &2i is given by (3)
terrains like the rural, open, urban and suburban.
suburban &2i = ( – )T (X−Xi) (3)
The Generalized Regression Neural Network (GRNN) Combining equations (1) and (2) and interchanging the
The GRNN was proposed by Donald F. Specht [12]. It is a order of integration and summation yields the desired
type of feed-forward neural network but not a standard conditional mean ().
backward propagation network. The GRNN is related to ./ 
∑ )
)1 ( *+,-  0
' 
the radial basis function neural network (RBF-NN)
(RBF and are σ
.
(4)
both classified under probabilistic neural networks
network (PNN). 
∑)1 *+,- / 0
σ
The GRNN is capable of approximating virtually any
function when given large amounts of data. Due to its
single-pass in nature, the GRNN requires a small fraction It is observed by [12]] that when the smoothing factor C is
of training data of backward algorithm to give the desired made large, the estimated density is forced to be smooth
output. As described in [1], the GRNN architecture consists and in the limit becomes a multivariate Gaussian with
of four layers:: input layer, pattern layer, summation and covariance σ2. On the other hand, a smaller value of C
output layer as shown in fig. 1. allows
ws the estimated density to assume non-Gaussian
non
shapes, but with the hazard that wild points may have too
great an effect on the estimate.

The Okumura Model

As described in [5], this
his is one of the widely used empirical
models for signal prediction. It was formulated by
Okumura based on drive test measurements
measurement in Japan at
operating frequencies ranging from 150 MHz to 1920MHz,
but extended to 3000MHz]]. The model is created for
macro-cellular networks that cover distances from 1km to
100km.. The height of base station antenna is between
30m to 100m. The propagation model is good in signal
prediction in terrains like urban, suburban, quasi-open
area and open areas.
Fig. 1: The GRNN Architecture
The model equation is given by (5)
L  LFSL 5 AMU – HMG – HBG –
GAREA
Input layer: This is the first layer that feeds into the
second layer called pattern layer. (5)

Pattern layer: The layer that is responsible for computing Where:

the Euclidean distance and activation function and later L = Median
ian path loss in Decibels (dB),
sends to summation layer which is the third layer.
layer LFSL = Free
ree Space Loss in Decibels (dB),
AMU = Median attenuation in Decibels (dB),
Summation layer: This layer is divided into two parts HBG = Base station antenna height gain factor given by
namely, numerator and denominator. The numerator
numer has 20log (hb/200) for 30m<hb<100m,
to do with the activities of summation of the multiplication HMG = Mobile station antenna
enna height
heig gain factor given by
of the training data and activation function. The 10log (hm/3) for hm<3m,
Denominator computes the activation function.
function GAREA = Gain due to type of environment

Output layer: This generates the output by dividing the The COST 231 Hata Model
numerator part by the denominator A described in [13], this model was created on the basis of
the Hata-Okumura
Okumura model by European cooperative of
The general equation as described by [12] [1 is as follows: Scientific and Technical
nical research, to suit European
given a vector variable, x, and a scalar variable, y, and terrains.. The model is widely used for signal prediction in
assuming X is a particular weighted value of the random wireless systems with minimum and maximum frequency
variable y, the regression of y on X is given by from 500MHz to 2000MHz. It has correction factors for
urban, sub urban, semi urban and rural areas. Due to its
E[y/x] =
∞ yf x, ydy/
∞ ,  (1)
∞ ∞ simplicity and availability of correction factors, it is widely
used for signal strength prediction [10]. The model
If the probability density function  (,)
expression is given by (6)
,) is unknown, it is
estimated from a sample of observations of x and y. The
probability estimator (,),
,), given by equation (2) is based (6)

@ IJTSRD | Unique Paper ID – IJTSRD30501

30501 | Volume – 4 | Issue – 3 | March-April
April 2020 Page 468
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
Where, Results and Analysis
L = Median path loss in Decibels (dB) It can be observed from Figures 2 to 7 that on the average,
C =0 for medium cities and suburban areas the empirical Okumura model overestimates the signal
C =3 for metropolitan areas strength across all Base Transceiver Stations. On the
f = Frequency of Transmission in Megahertz contrary, it can be observed that the COST 231 Hata
(MHz)(500MHz to 200MHz) significantly underestimates the signal strength.
hB = Base Station Transmitter height in Meters (30m to Furthermore, the figures indicate that, of the three
100m) predictors, the GRNN based model depicts the greatest
d = Distance between transmitter and receiver in prediction accuracy relative to the test data. The statistical
Kilometers (km) (up to 20kilometers) performance indices in Table 1 further buttress the fact
hm = Mobile Station Antenna effective height in Meters that on the average, the GRNN-based predictor is the most
(m) (1 to 10metres) accurate with a RMSE value of 5.8dBm and an R2 value of
a(hm) = Mobile station Antenna height correction factor as 0.71. It can be further observed that the Okumura model
described in the Hata Model for Urban Areas. with an RMSE value of 10.21dBm and R2 value of 0.168
outperforms the COST 231 Hata.
For urban areas, a(hm) = 3.20(log10(11.75hr))2−4.97, for f
> 400 MHz For sub-urban and rural areas, a(h R) =
(1.1log(f) - 0.7)h R - 1.56log(f) -0.8

Materials and Methods

A. Received Power Measurement
The terrain under investigation is the rural area between
the cities of Bauchi and Gombe, North-East Nigeria.
Received signal values were obtained from Base
Transceiver Stations (BTSs) belonging to the mobile
network service provider, popularly known as mobile
telecommunication network (MTN). The device used was
the Cellular Mobile Network Analyzer SAGEM OT 290. The
device is capable of measuring signal strength in decibel
milliwatts (dBm). Received power vaues (PR) were
recorded from the 1800 MHz frequency band at intervals
of 0.15 kilometer, after initial separation of 0.1 kilometer Fig. 2: Model Comparison for BTS 1
from the BTS.

The parameters obtained from mobile network service

provider include: Mean Transmitter Height (HT) of 40m
Mean Effective Isotropic Radiated Power (EIRP) of 46dBm.

B. Statistical Comparison Criteria

As described in [1], performance testing comparison of
model was based on two indecies,, namely, the Root Mean
Square Error (RMSE), and the Coefficient of Determination
(R2). The RMSE is a measure of the differences between
predicted and observed values. The RMSE expression is
given by (7)

L M
GHIJ  K∑N
$% (7)
N
Fig. 3: Model Comparison for BTS 2
Where,
M – Measured Path Loss
P – Predicted Path Loss
N- Number of paired values

R2 ranges between 0 and 1, but can be negative without a

constant, which indicates the model is inappropriate for
the data. A value closer to 1 indicates that a greater
correlation of the model with test data. The R2 is given by
(8)

∑S Q) QR) 
G  1 − ∑)1
S T ) 
(8)
)1Q) Q

where yi is the measured path loss, R$ is the predicted path

loss and T$ is the mean of the measured path loss.
Fig. 4: Model Comparison for BTS 3

@ IJTSRD | Unique Paper ID – IJTSRD30501 | Volume – 4 | Issue – 3 | March-April 2020 Page 469
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470

Fig. 5: Model Comparison for BTS 4 Fig. 6: Model Comparison for BTS 5

Fig. 7: Model Comparison for BTS 6

TABLE1. Statistical Performance Comparison of Predictors

MODEL STATS. BST 1 BST 2 BST 3 BST 4 BST 5 BST 6 MEAN
GRNN RMSE (dBm) 6.83 5.97 5.71 5.88 4.88 5.55 5.80
R2 0.59 0.76 0.73 0.63 0.80 0.74 0.71
RMSE (dBm) 7.68 9.71 11.40 9.92 11.20 11.37 10.21
Okumura
R2 0.51 0.40 1.01 0.18 0.18 0.16 0.16
COST RMSE (dBm) 16.75 13.76 11.69 12.90 12.17 11.50 13.13
231 Hata R2 -1.33 -0.20 -0.04 -0.40 0.03 0.14 -0.30

Conclusion [3] N. Rakesh, Dr. S. K. Srivatsa. A Sudy on Path Loss

This study considers the use of an artificial intelligence Analysis for GSM Mobile Network for Urban, Rural
based method, namely the GRNN, in mobile and Suburban Regions of Karnataka. International
communication network field strength prediction. The Journal of Distributed and Parallel Systems (IJDPS)
GRNN based predictor was created, validated and tested Vol.4, No.1, pp-53-66 2013.
with field strength data obtained from the terrain under
[4] Sunny Orike, Promise Elechi, and Iboro Asuquo
investigation. Performance comparisons indicate that
Ekanem. Assessment and Modeling of GSM Signal
GRNN based model proffers greater prediction accuracy
Propagation in Uyo, Nigeria. EJERS, European Journal
over the widely used empirical Okumura and COST 231
of Engineering Research and Science Vol. 2, No. 11,
Hata models.
pp-49-58, 2017.
REFERENCES: [5] Shoewu, O and F.O. Edeko. Analysis of radio wave
[1] Deme C. Abraham. Mobile Network Coverage propagation in Lagos environs. American Journal of
Determination at 900MHz for Abuja Rural Areas Scientific and Industrial Research. AJSIR ISSN: 2153-
using Artificial Neural Networks. International 649X, pp-438.455, 2011.
Journal of Trend in Scientific Research and
[6] A. Obot, O. Simeonb, J. Afolayanc. Comparative
Development (IJTSRD). e-ISSN: 2456 – 6470, Volume
Analysis of Path Loss Prediction Models for Urban
4 Issue 2, pp-1119-1123, 2020.
Macro Cellular Environments. Nigerian Journal of
[2] Deme A.C. Field Strength Determination in a Tropical Technology Vol. 30, No. 3, pp-51-59, 2011.
Metropolitan Environment Using Computational
[7] N. Y. Goshwe, J. Orinya, G. A. Igwue. Prediction of Path
Intelligence Techniques, 2016. American Journal of
Loss Estimate for a Frequency Modulation (FM)
Engineering Research (AJER) e-ISSN: 2320- 0847 p-
Station Using Basic Transmission Loss Model. Global
ISSN: 2320-0936 Volume-5, Issue-8, pp-276-284

@ IJTSRD | Unique Paper ID – IJTSRD30501 | Volume – 4 | Issue – 3 | March-April 2020 Page 470
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
Journal of Engineering Research Vol 12, ISSN 1596- European Journal of Engineering Research and
292x, pp- 69-75, 2013. Science Vol. 2, No. 11, pp-13-17, 2017.
[8] Abhayawardhana V. S., I. J. Wassel, D. Crosby, M. P. [11] Ogbulezie, J. C., Onuu, M. U., Bassey, D. E and
Sellers, M. G. Brown Comparison of empirical Etienam-Umoh, S. Site specific measurements and
propagation path loss models for fixed wireless propagation models for GSM in three cities in
access systems. 61th IEEE Technology Northern Nigeria. American Journal of Scientific and
Conference, Stockholm, 2005. Industrial Research. ISSN: 2153-649X, pp-238.245,
2013
[9] Callistus O. M, Joseph M. M, and Gabriel A. I.
Performance Evaluation of Generalized Regression [12] Specht D. F. A. A general regression neural network.
Neural Network Path Loss Prediction Model in IEEE Transactions on Neural Networks. 2, 1991, 568-
Macrocellular Environment. Journal of 576.
Multidisciplinary Engineering Science and
[13] Deme A. C. “Radio Propagation Modelling of a Typical
Technology (JMEST) ISSN: 3159-0040, 2015, pp.
Sudan Savanna Belt Rural Terrain using Soft-
2004 – 2008.
computing and Empirical Techniques”. Journal of
[10] Hammed Lasisi, Yinusa A. Adediran, and Anjolaoluwa Multidisciplinary Engineering Science and Technology
A. Ayodele. Development of Propagation Path Loss (JMEST) ISSN: 2458-9403 Vol. 6 Issue 12, pp 11320-
Prediction Model for Mobile Communications 11325, 2019
Network Deployment in Osogbo, Nigeria. EJERS,

@ IJTSRD | Unique Paper ID – IJTSRD30501 | Volume – 4 | Issue – 3 | March-April 2020 Page 471