See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/232895023

Traffic parameter estimation and highway classification: Rough patterns using a

neural networks approach

Article in Transportation Planning and Technology · February 1998

DOI: 10.1080/03081069808717607 · Source: CiteSeer

CITATIONS READS

6 522

1 author:

Pawan Lingras
Saint Mary's University
220 PUBLICATIONS 5,006 CITATIONS

SEE PROFILE

All content following this page was uploaded by Pawan Lingras on 21 May 2014.

The user has requested enhancement of the downloaded file.

 Rough Neural Networks

Pawan Lingras

Department of Computer Science, Algoma University College

Sault Ste. Marie, Ontario, Canada, P6A 2G4.
e-mail: [email protected]

Abstract

This paper describes rough neural networks which consists of a combination of rough
neurons and conventional neurons. Rough neurons use pairs of upper and lower bounds
as values for input and output. In some practical situations, it is preferable to develop
prediction models that use ranges as values for input and/or output variables. A need to
provide tolerance ranges is an example of such a situation. Inability to record precise
values of the variables is another situation where ranges of values must be used. In the
example used in this study, a number of input values are associated with a single value of
the output variable. Hence, it seems appropriate to represent the input values as ranges.
The predictions obtained using rough neural networks are significantly better than the
conventional neural network model.
 2

1. Introduction
The concept of upper and lower bound has been used in a variety of applications
in artificial intelligence (Shafer 1976; Pawlak 1982). In particular, theory of rough sets
(Pawlak, 1992, 1984) has demonstrated the usefulness of upper and lower bounds in rule
generation. Further developments in rough set theory (Polkowski, 1994; Wong, 1994;
Yao, et al, 1994), have shown that the general concept of upper and lower bounds provide
a wider framework that may be useful for different types of applications. This paper uses
rough patterns for predictions using neural networks. Each value in a rough pattern is a
pair of upper and lower bound. Conventional neural network models generally use a
precise input pattern in their estimations. The conventional neural network models need
to be modified to accommodate rough patterns. Rough neurons proposed in this paper
provide an ability to use rough patterns. Each rough neuron stores the upper and lower
bounds of the input and output values. Depending upon the nature of the application, two
rough neurons in the network can be connected to each other using either two or four
connections. A rough neuron can also be connected to a conventional neuron using two
connections. A rough neural network consists of a combination of rough and
conventional neurons connected each other. The paper outlines procedures for
feedforward and backpropagation in a rough neural network.
The paper also compares two different rough neural network models with a
conventional neural network model for prediction of the design hourly traffic volume
(DHV) for a highway section. The prediction is based on traffic volumes recorded over a
short period of time. The input to the network consists of traffic volumes for each day of
the week, i.e. Sunday, Monday, Tuesday, ..., Saturday, over the given time period. There
are several Mondays in the data collection period. Hence, the traffic volume for a Monday
cannot be a single value but must be a set of values. The conventional neural network
alternative uses the average of all the values for each Monday. Similar argument applies
to the rest of the days of the week. The use of average values tends to ignore some of the
available information. The rough neural network models use rough input pattern
consisting of upper and lower bounds of daily traffic volumes.
 3

2. Overview
This section briefly reviews some of the essential concepts of neural networks. A
brief description of highway data collection and analysis program is also provided.

2.1 Conventional Neural Networks

Neural networks are good at recognizing patterns, generalizing, and predicting
trends (White, 1989). Researchers have proposed different types of neural networks for
solving a variety of problems (Hecht-Nielsen, 1990). In its most general form, a neural
network consists of several layers of neurons. Each neuron receives input from other
neurons and external environment and produces output.
The conventional neural network used in this study is based on multi-layered, feed-
forward, and backpropagation design for supervised learning. This network consists of
one input layer, one output layer and one hidden layer of neurons. The input layer neurons
accept input from the external environment. The output from input layer neurons is fed to
the hidden layer neurons. The hidden layer neurons feed their output to the output layer
neurons which send their output to the external environment. Neurons from each layer
feed the output only to the next layer and hence the network is called feed forward. The
input and output of a neuron are governed by certain mathematical equations. Output
from a neuron is calculated using a transfer function.
Two stages in the development of the neural network model are training and
testing. During the training stage, the network uses an inductive learning principle to
learn from a set of examples called the training set. The learning process used in this
study is called supervised learning. In supervised learning, the desired output is known
for output layer neurons for the examples in the training set. The network attempts to
adjust weights of connections between neurons to produce the desired output. During this
process, the error in the output is propagated back from one layer to the previous layer by
adjusting weights of the connections. This is called the backpropagation method for
propagating the error. This study uses one of the most popular learning equations called
the generalized delta rule for supervised learning.
 4

In the testing stage, the network is tested for another set of examples for which the
output from the output layer neurons is known. After the neural net model is tested
successfully, it is used for predictions.

2.2 Estimation of Traffic Parameters

Highway agencies collect traffic volume data from various seasonal and
permanent traffic counters over a number of years (Garber and Hoel, 1988). Since the
installation of a permanent traffic counter (PTC) on every road section is not
economically feasible, highway agencies routinely use sample traffic counts (DeGarmo
and Sullivan, 1985). The sample traffic counts are obtained using seasonal traffic
counters (STCs). The data obtained from seasonal traffic counts is used to estimate
important traffic parameters for overall highway network (Sharma and Allipuram, 1993).
The present study deals with the estimation of an important traffic parameter called the
design hourly volume. A highway section is designed to service a certain amount of
traffic volume with a reasonable level of service. The volume that is used in such a design
is called the Design Hourly Volume (DHV). Different agencies use different criteria for
determining the DHV. However, most agencies use the highest hourly volumes in the
calculation of the DHV. The design hourly volume is calculated by sorting all the hourly
traffic volumes in a given year to identify the highest volume hours that are likely to
experience traffic congestion. The 30th highest hourly volume is one of the most
commonly used design hourly volumes.

s s s s s s

r r r r r r

(a) (b) (c)

Fig.1 Three Different Types of Connections Between Two Rough Neurons

 5

3. Rough Neural Network

A rough neural network consists of conventional neurons and rough neurons
connected to each other. A rough neuron r in rough neural networks can be viewed as a
pair of neurons, one for the upper bound called r and the other for the lower bound called
r. A rough neuron is connected to another rough neuron through two or four
connections. Fig. 1 depicts three types of connections between rough neurons. The
overlap between the upper and lower neurons indicates that upper and lower neurons
exchange information. Two rough neurons in Fig. 1(a) are fully connected. A rough
neuron r is said to be fully connected to another rough neuron s, if r and r are connected

to both s and s . If a rough neuron r is fully connected to s, then there are four
connections from r to s. In Fig. 1(b) and 1(c), there only two connections from r to s. If
the rough neuron r excites the activity of s (i.e. increase in the output of r will result in the
increase in the output of s), then r will be connected to s as shown in Fig. 1(b). On the
other hand, if r inhibits the activity of s (i.e. increase in the output of r corresponds to the
decrease in the output of s), then r will be connected to s as shown in Fig. 1(c).
This paper uses multi-layered, feed-forward, and backpropagation design outlined
in section 2.1 to describe the methodology of rough neural networks. Rough neural
networks used in this study consist of one input layer, one output layer and one hidden
layer of rough/conventional neurons. The input layer neurons accept input from the
external environment. The outputs from input layer neurons are fed to the hidden layer
neurons. The hidden layer neurons feed their output to the output layer neurons which
send their output to the external environment. The output of a rough neuron is a pair of
upper and lower bounds, while the output of a conventional neuron is a single value.
The input of a conventional, lower, or upper neuron is calculated using the
weighted sum as:

(1)
input i = ∑w ji× output j ,
there is a connection from j to i
 6

where i and j are either the conventional neurons or upper/lower neurons of a rough
neuron. The outputs of a rough neuron r is calculated using a transfer function as:

( )
output r = max transfer (input r ), transfer (input r ) , (2)

( )
output r = min transfer (input r ), transfer (input r ) . (3)

The output of a conventional neuron i is simply calculated as

output i = transfer (input i ) . (4)

This study uses the sigmoid transfer function given as:

1 (5)
transfer (u) = − gain× u ,
1+ e

where gain is a system parameter determined by the system designer to specify the slope
of the sigmoid function around input value of zero. There are several other functions for
determining the output from a neuron. The sigmoid transfer function is chosen because it
produces a continuous value in the 0 to 1 range.
If two rough neurons are partially connected, then the excitatory or inhibitory
nature of the connection is determined dynamically by polling the connection weights.
The network designer can make initial assumptions about the excitatory or inhibitory
nature of the connections. If a partial connection from a rough neuron r to another rough
neuron s is assumed to be excitatory and wr s < 0 and wr s < 0 , then the connection from

rough neuron r and s is changed from excitatory to inhibitory by assigning

wr s = wr s and wr s = wr s . The links (r , s) and (r , s) are disabled while links (r , s) and (r , s)

are enabled. On the other hand, if the neuron r is assumed to have an inhibitory partial
connection to s and wr s > 0 and wr s > 0 , then the connection between rough neuron r and
 7

s is changed from inihibitory to excitatory by assigning wr s = wr s and = wr s = wr s . The

links (r , s) and (r , s) are disabled while links (r , s) and (r , s) are enabled.

The training and testing stage in the development of a rough neural networks is
similar to the conventional neural network discussed in section 2.1. During the training
stage the network uses inductive learning principle to learn from a set of examples called
the training set. In supervised learning, the desired output from output layer neurons for
the examples in the training set is known. The network attempts to adjust weights of
connections between neurons to produce the desired output. During this process, the error
in the output is propagated back from one layer to the previous layer for adjusting weights
of the connections.
The weights of the connections are modified iteratively. The network is presented
with the training set repeatedly and is allowed to change weights after one (or more)
iteration(s). The weights are modified using a learning equation. This study uses the
generalized delta rule for modifying the weights of the connections using the following
equation:

w new ji + α × output j × errori × transfer ′(input i )

= w old
ji
(6)

where transfer ′( input i ) is the derivative of the transfer function evaluated at inputi and α
is the learning parameter which represents the speed of learning. For the sigmoid transfer
function used in this study,

transfer ′(input i ) = input i × (1 − input i ) (7)

The error in eq. (6) is calculated as:

errori = desired _ output i − output i . (8)

 8

As mentioned before, in the testing stage, the network is tested for another set of
examples for which the output from the output layer neurons is known. After the neural
net model is tested successfully, it is used for predictions.

4. Conventional and Rough Neural Network Models for Traffic Estimations

One of the objective of this study is to demonstrate the usefulness of rough neural
networks over the conventional neural networks. The neural networks have shown to be
more effective than the existing methods for estimation of traffic parameters such as the
Design Hourly Volume (DHV) for a highway section (Lingras and Adamo, 1995). This
section outlines the nature of the experiment and the rough and conventional neural
network models used in the study to predict the DHV.
Traffic volume data used in the study consisted of five year traffic volumes
collected at various permanent traffic counter (PTC) sites in the province of Alberta,
Canada. The PTC sites collect data for every hour in a given year. The annual hourly
patterns were divided into training and test sets. The hourly volumes for the entire year
were sorted and the thirtieth highest hourly volume was used as the DHV. For all the
objects in training and test sets, the DHV is known. The objective of the experiment is to
estimate the DHV based on daily volume patterns over a week in the months of July and
December. Months of July and December were chosen because these two months
generally have significantly different travel patterns.
Output Layer

Hidden Layer

Input Layer

Fig. 2. The conventional Neural Network Model Used in the Estimation of DHV
 9

Fig. 2 shows the conventional neural network model used for the estimation. The
conventional model has seven input neurons, four hidden layer neurons and one output
neuron. Neurons in the input layer are fully connected to neurons in the hidden layer.
Neurons in the hidden layer are fully connected to the neuron in the output layer. The
input to the conventional neural network model consists of average weekly pattern, i.e.
average daily volumes on Sundays, Mondays, Tuesdays, ..., Saturdays for an object. The
output is the DHV for the object.
The first rough neural network model shown in Fig. 3 has seven rough input
neurons, and eight hidden layer conventional neurons and one output neuron. Rough
neurons in the input layer are fully connected to conventional neurons in the hidden layer.
Conventional neurons in the hidden layer are fully connected to the conventional neuron
in the output layer. Since the hidden and output layer neurons are conventional neurons,
this network can be easily implemented using existing neural network packages such as
Stuttgart Neural Network Simulator (SNNS) (Zell, et al.).

Output Layer

Hidden Layer

Input Layer

Fig. 3. The Rough Neural Network Model Hidden Layer Conventional Neurons
 10

Output Layer

Hidden Layer

Input Layer

Fig. 4. The Rough Neural Network Model With Hidden Layer Rough Neurons

The second rough neural network model shown in Fig. 4 has seven rough input
neurons, and four hidden layer rough neurons and one output neuron. Rough neurons in
the input layer are fully connected to rough neurons in the hidden layer. Rough neurons in
the hidden layer are fully connected to the conventional neuron in the output layer. The
rough network shown in Fig. 4 can also be implemented using SNNS. However it was
necessary to add two activation functions to implement eq. (2) and eq. (3).
The input to both the rough neural network models consists of rough weekly
pattern, i.e. upper and lower bounds of daily volumes on Sundays, Mondays, Tuesdays,
..., Saturdays for an object. The output is the DHV for the object. Since the output is a
unique value, the output layer for both the models used a conventional neuron.
Both the networks were trained using 211 objects from the training set and tested
using 53 objects in the test set. Errors in estimation for the test set may originate from two
sources. One of the sources of errors is the sampling process. The number of patterns in
training and test sets might be very small, or the samples may not provide a good
representation of the universe. The other source of error is the estimation method itself.
In order to get an indication of the errors from these two different sources, the
conventional and rough neural networks were tested for training set as well as the test set.
Testing the models using the training set indicates how well the training method works by
itself.
The values of estimated and actual values of DHV are compared using the
following percent difference measure.
 11

estimated - actual
∆= × 100
actual

where
∆ = percent error
actual = actual DHV
estimated = estimated DHV
The maximum and average errors for each set are used to compare the results of
estimation. The average error provides a measure of the overall accuracy, while the
maximum error describes the worst case.

5. Result and Analysis

Table 1 shows the errors for the three models. Rough neural networks clearly
outperform the conventional neural networks. The reduction in errors from the
conventional network to the rough neural networks for training set is almost 50% for
average errors, while 75% for maximum errors. The reduction in errors for test set is 25%
for average errors and 10% for maximum errors. However, sampling errors may explain
somewhat lower performance gain.

Model Train Set Test Set

Maximum Average Maximum Average

Conventional 46.2% 9.6% 28.1% 9.7%

Rough 1 17.5% 5.5% 24.9% 8.1%
Rough 2 13.7% 5.8% 23.0% 8.0%

Table 1. Percetage Errors for Conventional and Rough Set Models

 12

1.5 Conventional Model

Errors

Rough Model 2 Rough Model 1

0.5

0
0 5000 10000 15000 20000 25000
Iterations
Fig. 5 Sum of Squares of Errors for Conventional and Rough Neural Networks

The first rough neural network with conventional neurons in the hidden layer
results in somewhat higher errors than the second rough neural network which uses rough
neurons in the hidden layer. This indicates that use of rough neurons in place of
conventional neurons where ever possible will improve the performance of the neural
networks. Another interesting observation from the experiment is the change in the sum
of squared errors during the training process. Fig. 5 shows reduction of errors during the
training process for all the three networks. The reduction in errors is more dramatic for
rough neural networks than the conventional network. The errors for second rough neural
network are consistently lower than the first neural network confirming earlier
observation regarding the use of hidden layer rough neurons.
 13

6. Summary and Conclusion

This study proposed rough neural networks for estimating rough output patterns
from rough input patterns. A rough pattern uses upper and lower bounds of the values as
opposed to precise values. The rough neural networks use a combination of rough and
conventional neurons. A rough neuron can be viewed as a pair of neurons. One neuron
corresponds to the upper bound and the other corresponds to the lower bound. Upper and
lower neuron exchange information with each other during the calculation of their
outputs. The paper discussed different types of connections to and from rough neurons
and the corresponding feedforward mechanisms.
The study compared the effectiveness of two different rough neural network
models with a conventional neural network model for estimation of the design hourly
traffic volume on a highway section. The problem lends itself very well for rough neural
network modeling. The errors in estimation from rough neural network models are
significantly lower than the conventional neural network model. Moreover, the addition
of rough neurons in hidden layer seems to improve the prediction performance.

References
DeGarmo, E.P., Sullivan, W.G. and Canada, J.R. 1984. Engineering Economy,
Macmillan Publishing Co., New York, N.Y., pp. 264-266.
Garber, N.J. and Hoel, L.A. 1988. Traffic and Highway Engineering, West Publishing
Co., New York, N.Y., pp. 97-118.
Hecht-Nielsen, R. Neurocomputing, Addison-Wesley Publishing, Don Mills, Ontario.
Lingras, P.J. and Adamo, M. (1995). Estimation of AADT Volume Using Neural
Networks, Computing in Civil and Building Engineering, Pahl & Werner (Eds.),
1355-1362.
Pawlak, Z. (1982). Rough sets, International Journal of Information and Computer
Sciences, 11, pp. 145-172.
Pawlak, Z. (1984). Rough classification, International Journal of Man-Machine Studies,
20, pp. 469-483.
 14

Polkowski, L., Skowron, A. and Zytkow, J. (1994). Rough Foundations for Rough Sets,
Conference Proceeding of Third International Workshop on Rough Sets and Soft
Computing, November 10-12, San Jose, California, pp. 142-149.
Shafer, G. (1976). A Mathematical Theory of Evidence, Princeton University Press,
Princeton, New Jersey.
Sharma S.C. and Allipuram, R.R. 1993. Duration and Frequency of Seasonal Traffic
Counts, Journal of Transportation Engineering, American Society of Civil
Engineers, 116, 3, pp. 344-359.
Sharma S.C. and Werner, A. 1991. Improved Method of Grouping Provincewide
Permanent Traffic Counters, Transportation Research Record 815, Transportation
Research Board, Washington D.C., pp. 13-18.
White, H. 1989. Neural Network Learning and Statistics, AI Expert, 4, 12, 48-52 .
Wong, S.K.M. (1994). Rough Sets and Extended Models, invited paper in the Third
International Workshop on Rough Sets and Soft Computing, November 10-12, San
Jose, California.
Yao, Y.Y., Li X., Lin, T.Y. and Liu, Q. (1994). Representation and Classification of
Rough Set Models, Conference Proceeding of Third International Workshop on
Rough Sets and Soft Computing, November 10-12, San Jose, California,
pp. 630-637.
Zell, et al. (1995). Stuttgart Neural Network Simulator: User Manual Version 4.0,
University of Stuttgart, Institute of Parallel and Distributed High Performance
Systems, Report No. 6/95.

View publication stats