A Discrete Markov Model of Vehicular Traffic

Author(s): MICHAEL C. DUNNE, RICHARD W. ROTHERY and RENFREY B. POTTS

Source: Transportation Science, Vol. 2, No. 3 (August 1968), pp. 233-251
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/25767487
Accessed: 08-12-2016 04:54 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Transportation Science

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 A Discrete Markov Model of Vehicular
Traffic*
MICHAEL C. DUNNEf and RICHARD W. ROTHERY

General Motors Research Laboratories, Warren, Michigan

and

RENFREY B. POTTS*

University of California, Berkeley, California

For a number of years the Poisson process has been used as a model
to describe the time series of events of vehicles passing a point on
a roadway. In its simplest form the Poisson model does not take into
account the finite size of the vehicles nor does it properly describe
the characteristic bunching of traffic that commonly occurs, for
example, on a two-lane road when passing is hindered. To give a
more comprehensive description of traffic, various extensions of the
Poisson model and a variety of other models have been proposed,
with varying success. In this paper a discrete Markov model of
traffic is proposed. The model is an extension of the simple bi
nomial model and accounts for the bunching tendency of traffic by
assuming correlations between successive vehicles. A description
is given of the time-arrival experiments that were carried out to
test the model and the statistical analysis of the results of these
experiments is discussed. In addition, several theoretical aspects
of the model are examined.

* An account of this work was presented at the Transportation Science Section

Session of the Thirty-second National meeting of the Operations Research Society
of America, Chicago, November 1-3,1967.
t Present address: Institute of Highway and Traffic Research, The University of
New South Wales, Australia.
t Permanent address: The University of Adelaide, South Australia.
233

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
234 / DUNNE, ROTHERY, AND POTTS

In recent years many mathematical models of vehicular traffic have been

proposed,[1"101 and the most useful have been the simplest?the fluid model
in which the traffic stream is regarded as a continuous fluid flowing at a
uniform rate, and the Poisson model in which the traffic is assumed to be
random. The fluid model has obvious limitations but the Poisson model
has been successfully applied to a wide variety of situations involving light
traffic and traffic in which passing is unhindered.171
Stochastic models of traffic are usually described by the headway and
counting distributions that can be determined by a time-arrival experiment
in which the instances when vehicles (their front wheels, for example)
pass a fixed point on the road are recorded. From these data, the head
ways or time intervals between successive vehicles can be calculated as
well as the counts or number of cars passing the point in a given time
interval (or time slice). For the Poisson model, the arrival times form a
random series of events, the counting distribution is Poisson and the head
ways are independently and identically distributed according to the ex
ponential distribution.
There are two characteristics of vehicular traffic that distinguish it
from, say, telephone traffic and that are not exhibited by the simple Poisson
model. The first characteristic is the discrete microscopic structure of the
traffic, consisting, as it does, of individual vehicles with finite lengths.
This structure prevents the occurrence of very small headways in single
lane traffic?yet the Poisson model assigns probabilities that increase in
magnitude to a maximum as the headways decrease to zero. The other
characteristic is the tendency of traffic to bunch when passing is hindered,
as on a road with two-way traffic. For the Poisson model the coefficient of
variation, C, of the headways and the variance to mean ratio, /, of the
counts are both unity whereas bunching of traffic is indicated by experi
mental values as high as two for these quantities.
To overcome the limitations of the Poisson model, several more compli
cated traffic models have been proposed. To prevent small headways,
shifted headway distributions have been suggested, and bunching has been
introduced in branching Poisson and semi-Markov models.141 It is the pur
pose of the present paper to investigate the properties of a two-state Markov
model that is a generalization of the simple binomial model.C81 Unlike the
binomial model, the Markov model accounts for bunching of traffic by
introducing correlations between successive vehicles.
To test a traffic model from data collected by a time-arrival experiment
requires considerable care. On the one hand, a reliable statistical analysis
requires a long run of data free from systematic trends, yet traffic conditions
are likely to vary even over fairly short periods of time. The common de

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 a discrete markov model of vehicular traffic / 235

cisions to choose time slices of 60, 30, 15, 5, 1 or 3^ minute durations are
rather arbitrary and it is well known that different time slices lead to quite
different conclusions.17?91 Cox and Lewis m have suggested a procedure,
that is followed in this paper, for the thorough statistical analysis of a series
of events and used as an example some Swedish traffic data that have been
previously analyzed by Miller[10j and Bartlett.[4J The coefficient of
variation of the headway distribution for this data was 1.5 and tests
proved that the Poisson model was inadequate, but that a satisfactory
fit could be obtained with the branching Poisson and semi-Markov models
although there was insufficient data to discriminate between them. It is
the purpose of this paper to test the adequacy of the two-sta,te Markov
model by applying the methods proposed by Cox and Lewis to time-arrival
data collected for traffic on one lane of a two-way road.

TWO-STATE MARKOV MODEL

Consider one lane of a two-lane road (one lane in each direction). Sup
pose that the time scale is divided into consecutive nonoverlapping inter
vals of length r sec?i.e., intervals of the form (fcr, fcr + r), fc = 0,1,2, .
Let

Xk = number of vehicles that pass a fixed point on the road- ^

way in (fcr, fcr + r).
The possibility of two or more vehicles passing the same point in a r-sec
interval is ignored so that Xk = 0 or 1 for all fc. In this paper, in contrast
to the simple binomial model wherein the Xks are assumed independent,
it will be assumed that traffic can be described by a Markov process with
stochastic matrix

(2)
with 0^p^l,0^^1 and p and q independent of fc. Hence, for
example,
P[Xk+1 = 11 Xk = 0] = p, (3)
P[Xk+l = 11 X* = 1] = 1 - g, (4)
and so on. When p = 1 ? q this Markov model reduces to the simple
binomial model.
There are two fundamental properties of the Markov model that are of
great importance and that will be referred to in later sections. First,
like the random arrivals model, it is a renewal model, i.e., the times be
tween events (arrivals) are independently and identically distributed.

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 236 / dunne, rothery, and potts

Second, by its very definition, the model implies the existence of corre
lations between the counts of vehicles in successive time intervals and this
property is not possessed by the random arrivals model.

Counts
The counting distribution is characterized by

Yn = Zi + X2 + + Xn . (5)
The properties of Yn are well known 111'141?for instance Yn is asym
cally normal with mean np/(p + q) and variance npq(2 ? p ? q)/(p
Hence the traffic flow is Q = 3600p/{(p + q)r] veh/hr and the va
to mean ratio of counts, 7, tends to q(2 ? p ? q)/(p + q)2.

Platoons and Gaps

For the purpose of this paper a platoon of length k will be defined
consecutive Ts' and similarly a gap of length k will be defined as k c
secutive 'OV. Both platoons and gaps have the geometric distribution
A summary of the properties of the two state Markov model is giv
the Appendix.

EXPERIMENTAL DATA

Data were taken on one lane of a two-lane rural road at a point 2.4 mi
downstream from the nearest stop sign or traffic signal. The passage of
each vehicle traveling in a single lane was recorded on magnetic tape to
gether with a 200 cycle/sec clock signal using standard digital recording
techniques. This information was then transcribed from magnetic tape to a
digital format suitable for computer analysis.
These experiments were conducted between 3:00 pm and 6:00 pm on
several days over a period of approximately two months. Flow levels from
about 300 to 650 vph were observed. The duration of each data set ranged
from one-half to one- and one-quarter hours depending on traffic conditions.
Two additional sets of data were collected at a second location (on a differ
ent road) to exclude the possibility that the results might have been biased
by features peculiar to the first location. These data have been included
in the analysis and possess the same characteristics as the others.

STATISTICAL ANALYSIS OF DATA

As the parameters (p> q, and r) in the Markov model are time independent,
no allowance is made for variations in flow level. For this reason a meaning
ful analysis of a data set can only be made when the flow remains reason

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 A DISCRETE MARKOV MODEL OF VEHICULAR TRAFFIC / 237

ably constant over the period of observation. One of the simplest, yet most
revealing, methods of checking for systematic trends in the flow level is to
plot the cumulative number of arrivals, N(t), as a function of t. Such a plot
reveals how the average flow fluctuates with time and enables a trend-free
slab of data to be found graphically by use of a straightedge.
Fig. 1 illustrates how data set No. 1 has been divided into three sections
of approximately constant flow. Out of all the data collected, nine trend
free sections were obtained graphically and were subject to further analysis.
As indicated in Table I the lengths of the trend-free sections varied from

400 r

Msec.)

Fig. 1. Plot of the cumulative number, N(t), of arrivals up to time

t against t for data set No. 1.

about 10 to 20 min and the flow levels ranged from 350 to 650 veh/hr.
In all cases the more sophisticated tests described by Cox and Lewis15'121
when applied to the original data sets, indicated the existence of systematic
trend and confirmed that the sections selected for further analysis were
trend free.

Testing the Renewal Hypothesis

The acceptance of the renewal hypothesis requires that the interarrival
times be independent. For this purpose it is necessary to obtain estimates
of the serial correlation coefficients that characterize the second-order
joint properties of the interarrival times. Let n be the number of inter

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
to O

3.224**,
3.677**, 3.192**, 1.233, -0.589,
-1.747,
5.691**,
2.587**,
-0.835,0.098,0.2911.602
i (2).i - 1. 2.

-2.401 -1.016,
4.234**,
1.178 2.178*, -0.564
0.614, -0.076,0.804
-0.321,
2.358 -0.540
1.694,
0.168, 0.992, -0.237 0.123, 1.579,
-0.9131. 5 19, 0 . 4 71, 0 . 0 46 3.771**, -0.530

1, 2, ,9 -0.803, -0.159, -1.471 -0.955, -0.585, -1.545 -0.153, -0.564, 0.240

-0.502, -0.430, -0.208-1.192, -0.489,-0.675

-1.111, -1.033
0.613, -1.272, -0.539

0.855, -0.035,
0.398, -0.241
-0.564, -1.132 -0.578,-1.156,0.841,
0.662, 0.1261.588, -0.583
0.052, 0.079, 0.110

Dn
TABLE I

1.742** = Dn 2.738** = Dn
2.197**
9.235** 5.030** 11.701** 2.257** 8.249**
0.993 0.594

Distribution-Free1.948
Statistics 1.406
(Transformed Data)

Dn+Dn~Wn2 Dn+Dn~Wn2 Dn+Dn"Wn2 Dn+D^Wn2

Dn
172.8**
0.907 = Dn
Data) 0.705
0.473
0.731176.6 0.901 0.878 0.517

0.491 0.84499.7 0.776 0.562

Distribution-Free

Statistics (Untransformed

= 1.224128.9**
L DrT In In 2>n+
In
Dn+DnWn2 Dn+ Wn2 Dn+Dn~Wn2 DrTWn2

Q = 592 veh/hr Q = 356 veh/hr

Data Set Character

istics Q = 381 veh/hr
C = 1.376

Q = 585 veh/hr
1168 sec
C = 1.190 669 sec C = 1.126 1227 sec
130 veh C = 1.344 88 veh
890 sec

190 veh 110 veh

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 o< o O w to
CoCD

-1.954, -0.704, -1.412 -0.663, -1.146, -1.633

-0.324, -0.423,
-1.106,-0.282,
0.2740.741
0.841,
4.144**, 3.974**, -0.755, -1.354
2.236
3.682**, 1.441, 0.910 2.572**, 1.091, -0.558 3.694**, 2.432*, 2.948
0.621, 0.437, -2.689** 1.747, 2.115*, 2.101* 1.966*, 1.047, -0.067
0.804, 0.461, 1.788 1.682, 1.202, -1.177

0.378, 0.610, -0.308

-0.729, -1.231, -0.032

-1.862, -1.304, 0.480 -0.774, -0.661, 1.684-1.364, -0.468, -0.770 -1.204, -0.405, -0.843 -0.644, -1.106, 0.051

-0.027,
-1.145, -0.245, -1.077,
-0.727 0.347
0.137, -1.116,-0.203,
-1.528 2.440*, -0.252 -0.685, 0.711, -0.035 1.007, -1.358, -0.373

-0.674, 0.415, -0.893

0.123, 1.702, -1.278

1.815, 1.327, 0.256

= Dn

Dn+ = 2.096** =? Dn
Dn+ = 1.789** = Dn 2)n- = 2.092** = D? Dn+ = 1.425* ? Z>?

1.488*1.3445.219*
Wn2 = 8.660**

Dn" = 1.448
JF?2 = 6.630** Dn+ =1.287 TFn2 = 6.184** Dn" = 2.087 AT = 1.112JF?2 = 3.569**

Dn+Dn"Wn2
2>?

Dn+ = 0.637 = Z>?

2>n+ = 0.488 = Dn Dn~ = 0.794 = Dn Dn+ = 0.678 = Dn

Dn~ =Wn2
0.444 = 0.586
In = 132.4 Zn - 134.9
AT = 0.479 Wn2 = 0.633 ln = 117.8 Z? = 183.8
Wn2 = 0.453 Dn" Wn2 Z? = 105.0
= 0.660= 0.554

2)n+ = 0.673
/)?- = 0.711
TFn2 = 0.634 Z>?+ = 0.326

Q = 654 veh/hr Q = 528 veh/hr Q = 579 veh/hr Q = 388 veh/hr

C = 1.189 C = 1.154

Q = 524 veh/hr
C = 1.316 C = 1.122

887 sec
129 veh 826 sec 818 sec
1181 sec
900 sec C = 1.137
150 veh 120 veh 190 veh 97 veh

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 240 / DUNNE, ROTHERY, AND POTTS

arrival times and pj the estimates of the serial correlation coefficients

pj where j = 0, 1, 2, is the lag index, i.e., refers to interarrival times
separated by j ? 1 consecutive interarrival times. Then, under the null
hypothesis py = 0, j = 1, 2, the quantities f>/ = (n ? j)1/2pj are
approximately normally distributed with mean 0 and variance 1. Hence,
testing the renewal hypothesis is equivalent to testing the null hypothesis
= 0.

4.0 r

3.0 h

2.0

1.0 -

; J_j_
"1' 111111
-1.0 -

-2.0

-3.01 J

-4.0 L
Fig. 2. Estimates of the first 50 standardized serial correlation coefficients for the
interarrival times for data set No. 1. The horizontal dashed lines are 5 per cent sig
nificance bands and the solid lines are 1 per cent significance bands for the estimates.

The values of the !>/ for j = 1, 2, , 9 are shown in Table I to indicate

their order of magnitude. An examination of these quantities for higher
values of j leads to the conclusion that all data sets with the exception of
5 and 6 are consistent with the renewal hypothesis. For the data sets that
are not consistent with the renewal hypothesis it is hard to comprehend
how a particular interarrival time might be correlated with another one
separated from it by 10, 20, or 30 interarrival times (i.e., separated by as
much as 3 min!).
Fig. 2 shows l>y forj = 1, 2, , 50 for one of the data sets that is con
sistent with the renewal hypothesis.

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 A DISCRETE MARKOV MODEL OF VEHICULAR TRAFFIC / 241

Testing for Randomness

A preHminary indication of the adequacy, or otherwise, of the Poisson
hypothesis is obtained from the values of the coefficient of variation of the
headway distribution and the value of the variance to mean ratio of the
counts. The latter depends, quite critically, on the time slice so it is im
portant to use more sensitive tests in deciding upon the acceptance or
rejection of the Poisson hypothesis. The tests used are uniform conditional
tests and, for their application, four distribution-free statistics151 are
computed, viz., the onesided Kolmogorov-Smirnov statistics, Dn+ and
Dn", the two-sided Kolmogorov-Smirnov statistic, Dn = max (Dn+, DrT),
and the Anderson-Darling statistic Wn2. In addition the Durbin trans
formation1131 has been applied to the original data and another set of values
of the distribution-free statistics obtained. The reason for using this trans
formation is that there is considerable evidence that tests based on the
transformed data have much greater power against a wide range of alterna
tives. Finally the Moran statistic, ln , has been calculated.
The values of the distribution-free statistics computed to test the
Poisson hypothesis are shown in two columns of Table I. A double asterisk
indicates rejection of the hypothesis at a level less than the 1 per cent level
while a single asterisk indicates rejection at a level between the 1 per cent
and 5 per cent levels. Those statistics calculated from the untransformed
data are, with a couple of exceptions, not significant whereas those based
on the transformed data lead to rejection of the Poisson hypothesis.

Correlation between Counts

Correlation between the counts in r-sec intervals is a direct consequence
of the Markov property of the model. Tests for the existence of these cor
relations are entirely analogous to those used in testing the renewal hypothe
sis (i.e., in testing whether the interarrival times are correlated). In par
ticular, if m is the number of r-sec intervals, c,(r) the covariance between
counts in two r-sec intervals separated by j ? 1 r-sec intervals, ixj(r) the
corresponding serial correlation coefficients and fij(r) their estimates!
then, under the null hypothesis jxj{r) = 0, j = 1, 2, the quantities
Bj(r) = (m ? y)1/2/Iy(r) are approximately normally distributed with
mean 0 and variance 1. Hence the rejection of the null hypothesis will
imply the presence of correlation between vehicle counts.
The values of 0y(r), j = 1, 2, , 9 are given in Table I for r = 2 sec.
Clearly, at least for this value of r, the null hypothesis Hj(r) = 0 is re
jected, i.e., there is correlation between the counts in 2-sec intervals. In
t The formulas used for the estimates pj and ?;(t) are those given in Chap. 5 of
reference 5.

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
242 / DUNNE, ROTHERY, AND POTTS

fact, in all but one case, the level of significance is less than 1 per cent.
The correlations are strongest iorj = 1 or 2 indicating 'nearest neighbor'
or 'next-nearest neighbor' correlations?in other words a '1' is likely to be
followed by a '1' in the next interval or in the next but one.
Although the 0y(r) are dependent on r, the existence of correlations is a
physical property of the traffic stream and if correlations exist for some
particular value of r they will exist for all other sensible values of r, but
may be of different sign and magnitude. For example, Fig. 3 shows 0;(2) for

r= 2.0

-4.0 L
Fig. 3. Estimates of the first 50 standardized serial correlation coefficients for the
counts in 2-sec intervals for data set No. 1. The horizontal dashed lines are 5 per cent
significance bands and the solid lines are 1 per cent significance bands for the esti
mates.

j = 1, 2, , 50 for a particular data set and Fig. 4 shows 0/(0.5) for

j = 1, 2, , 50 for the same set of data. In both cases correlations exist
but, while 0i(2) and 02(2) are positive, 0i(O.5) is negative, and 04(O.5) and
?s(0.5) are positive (by 'negative' and 'positive' are meant significantly
negative and significantly positive). That 0i(O.5) is negative is not sur
prising as it indicates that there are relatively few headways in the range
of 0.5 sec to 1 sec (i.e., a '1' is likely to be followed by a '0'). Also, knowing
that 0i(2) is positive it is to be expected that 04(O.5) would be positive as

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 A DISCRETE MARKOV MODEL OF VEHICULAR TRAFFIC / 243

both quantities measure correlations between vehicles separated, on the

average, by 2 sec.
The existence of correlations in the traffic streams is thus strongly im
plied. It is, however, very difficult to say anything much more precise
about these correlations than that they do exist.

T- 0.5

nil,II _LL

-4.0 L
Fig. 4. Estimates of the first 50 standardized serial correlation coefficients for the
counts in 0.5-sec intervals for data set No. 1. The horizontal dashed lines are 5 per
cent significance bands and the solid lines are 1 per cent significance bands for the
estimates.

Model Calibration
The parameters to be estimated are p, q, and r, p and q being dependent
on r. For a given set of arrival times there is associated with each value of
t a unique binary sequence from which p and q can be calculated in a
straightforward manner. Choosing a suitable value of r is, however, a more
difficult and far less well defined procedure. As this procedure is based on
the behavior of the serial correlation coefficients, hj(t), defined hitherto,
some of their properties will now be mentioned. In what follows the argu
ment, r, will be omitted for simplicity of notation.
It can be shown that jij = ju0(l ? p ? qY where /x0 is positive. Hence

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
244 / DUNNE, ROTHERY, AND POTTS

when p + q < 1, nj > 0 for all j and ah > /*2 > while for p + q > 1,
H2j > 0 and M2y+i < 0 for all j and | m \ > M2 > . As has been noted
before, when p + q = 1 the Markov model reduces to the simple binomial
model for which m = 0 for all j. The quantities 0/ have exactly the same
properties of sign and order.
The fact that it is assumed that no more than one vehicle should arrive
in each r-sec interval imposes an upper bound on the size of r. This con

8.o r

-6.0 -

-8.0 L
Fig. 5. Plot of the first 3 standardized serial correlation coefficients for the counts
in r-sec intervals against t for data set No. 1. The horizontal dashed lines are 5 per
cent significance bands and the solid lines are 1 per cent significance bands for the
estimates.

dition is, however, too restrictive to be practicable for, even under condi
tions of light flow, there will always be a few very small headways (say, of
the order of 3^ sec) that may lead to the inclusion of 2 veh within a r-sec
interval unless t is chosen smaller than the smallest headway. In addition,
choosing r to be of the order of ^ sec leads also to inconsistencies in the
properties of the serial correlation coefficients. So, rather than insisting
that no 2's should occur, it is suggested that r be chosen so as to keep the
number of 2's insignificant.
In Fig. 5 a plot is shown of 6j against r for j = 1, 2, 3. In the range
0,5 ^ t ^ 0.8, 0i is negative and 03 is positive, which is inconsistent with

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 A DISCRETE MARKOV MODEL OF VEHICULAR TRAFFIC / 245

the properties of the Bj. For values of x ranging from 0.8 up to about
1.6, 02 is greater than 0i, which, again, is not acceptable. For values of r
larger than about 1.6, the Bj behave in accordance with the properties of
the Bj.
TABLE II
p q Number of 2's

1.0 0.17 0.88 1

1.5 0.22 0.72 4
2.0 0.25 0.58 13
2.5 0.28 0.50 23

Frequency

A?/A
/ / \

10 12
Number of Vehicles

Fig. 6. Comparison of theoretical counting distribution with that obtained ex

perimentally for data set No. 1. The theoretical distribution is represented by the
solid curve and the experimental distribution by the dashed lines. The time slice is
T = 30 sec and the experimental values of the mean and variance are 4.54 and 5.58
respectively.

In Table II values of p and q, together with the number of r-intervals

with 2 veh arrivals, are given for r = 1.0, 1.5, 2.0, and 2.5 sec. Certainly
there is no hesitation in rejecting a value of r = 2.5 because of the high
number of 2's associated with this value. In fact r = 2 produces about as
many 2's as can reasonably be tolerated.

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
246 / DUNNE, ROTHERY, AND POTTS

From these considerations one would therefore choose a value of r in the

approximate range 1.6 to 2.0. It is not possible to pick any exact value of
t from the above reasoning and, indeed, one would not expect to as the
model would hardly be sensible if it could be said that r = 1.8 was the
correct value of r while r = 1.7 or r = 1.9 were incorrect values.
There is no point in examining the 0/ for very small values of t as it is
obvious for physical reasons that they will all become negative?as r de

70i

Frequency

Platoon Length

Fig. 7. Comparison of theoretical platoon length distribution with that obtained

experimentally for data set No. 1. The theoretical distribution is represented by the
solid curve and the experimental distribution by the dashed lines. A value of 2 sec
has been used for t.

creases, first 0i becomes negative, then 02, and so on. In the given example,
0i becomes negative at r = 0.8 or 0.9 and 02 appears certain to become
negative at about r = 0.4. No further information is to be gained by in
vestigating the behavior of the 0/ for values of j greater than 3 since, except
again for the very small values of r, these will quite quickly tend to zero.
Once a value of r is chosen the values of p and q are calculable merely
by counting the number of Ts following l's and so on.
In summary: the range of values of r is found for which the 0/ exhibit the

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 A DISCRETE MARKOV MODEL OP VEHICULAR TRAFFIC / 247

same properties of sign and magnitude as do the 0, of which they are

estimates. Then an upper bound on t is fixed by the requirement that
'not many' 2's should appear. The parameters in the transition matrix are
then easily found.

Macroscopic Properties
The macroscopic traffic properties predicted by the model can be de
scribed by the frequency distributions for vehicle counts, and platoon and

Frequency

Gap Length

Fig, 8, Comparison of theoretical gap length distribution with that obtained

experimentally for data set No. 1. The theoretical distribution is represented by the
solid curve and the experimental distribution by the dashed lines. A value of 2 sec
has been used for t.

gap lengths. For data set No. 1, with r = 2, the transition matrix is
).75
p =. r?-7 0.25]
).58 0.42J

and in Fig. 6 the corresponding counting distribution is compared with that

obtained from the experimental observations. Comparisons between the
theoretical and experimental platoon and gap distributions are shown in
Figs. 7 and 8.

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
248 / dunne, rothery, and potts
DISCUSSION

The present paper has considered the properties of a two-state Marko

model of traffic and has tested the adequacy of the model for describing
the arrivals of vehicles at a point on a roadway when passing is hindered
and the traffic flow is medium to heavy. Under such traffic conditions th
commonly used Poisson model is inadequate because it fails to account fo
the bunching of traffic that, in the two-state model, is a consequence of
postulatedintervehicle correlations.
The two-state model includes as a special case the binomial model an
shares with it desirable discrete properties that in a sense reflect the true
vehicular nature of a traffic stream. Traffic platoons and gaps are defined
in lerms of the presence or absence of a vehicle in each of a succession o
discrete r-intervals. In order to derive explicit expressions for headway
distributions, it could be assumed, for example, that the distribution of
arrivals within each r-sec is uniform, an assumption made by Allan t6] i
his discussion of the continuous binomial model.
A statistical analysis of traffic data collected specifically for this project
has highlighted some of the major difficulties inherent in testing traffi
models. Macroscopic properties that can be described by first and second
moments of the headway and counting distributions are insensitive to th
microscopic model details, and the effects of different choices of time slice
may foil attempts to draw reliable conclusions from traffic data. By way o
illustration, one 30-min section of data from traffic at a flow of about 60
veh/hr was divided into time slices of 90 sec, and the following results fo
the mean, variance, and variance-to-mean ratio for the vehicle counts wer
obtained using three adjacent time origins U = 0, ?0 = 30, t0 = 60 sec:
t0 = 0 to = 30 to = 60

mean 14.6 14.7 14.7

variance 17.4 19.3 37.3
variance/mean 1.19 1.32 2.54

This example illustrates how sensitive the variance can be to a sma

in the time origin. This effect, together with the influence of the mag
of the time slice itself, makes it impossible to draw reliable conclu
from these data about the second-moment properties of the count
tribution. To increase the reliability, more data are required b
difficult to obtain long traffic runs without a trend in the flow. It
seem that these difficulties in analyzing traffic data have not been p
appreciated.
In the present paper, the sophisticated statistical techniques described

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 a discrete markov model of vehicular traffic / 249

by Cox and Lewis have been used to establish the adequacy of the renewal
hypothesis, to reject the random arrivals model, and to prove the existence
of intervehicle correlations. However, it is not claimed that the two-state
Markov model is in any sense 'the' model of traffic. The more modest claim
is that the model can adequately describe the bunching of traffic that is an
important characteristic of medium traffic when passing is hindered. The
model is simple in form and directly describes the microscopic features of
traffic and infers the macroscopic properties, rather than obtaining them
from curve fitting techniques.
Possible extensions of the model that are being investigated include a
3-state model, an analysis of pulsed traffic and its diffusion, and a model of
correlated platoons and gaps.

APPENDIX
Two-State Markov Model: Some Mathematical Properties

The mathematical properties of this model have appeared elsewhere (see, for
example, reference 11) but for completeness some of the more important ones are
included here.

Transition Probabilities
The elements of the matrix Pn are the n-step transition probabilities. The simplest
method of evaluating Pn is that of spectral resolution for which purpose the eigen
values of P must be found. They are 1 and 1 ?p ? q respectively and the required
form of Pn is

p] + d - p - g)n T p -pi
v+qL
?

v+qL

Counting Distribution
To determine the probability generating function of the number, Yn, of suc
cesses in n trials it is necessary to evaluate [P(z)]n where

The eigenvalues Ai(z), \(z) of this matrix are solutions of the quadratic equation

X2 - A[l - p + (1 - q)z] + (1 - p - q)z = 0.

The method of spectral resolution can be used to give [P(z)}w but the implicit form
is omitted.

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
250 / dunne, rothery, and potts

If the mean and variance of yn are denoted by e(yn), v(yn), Then

e(yn) = np/(p + q),
v(yn) = npq(2 - p - q)/(p + q)* - 2pq(\ - p - q) [1 - (1 - p - q?/(p + <?)4
^ wpg(2 ? p ? q)/(p + qY.
It can be shown by a central limit argument that yn is asymptotically normal.

Correlation between Counts

The serial correlation coefficients pj [strictly, pj(t)] have the form
pq n \
ft = (M^(1-p-?)'
Platoons and Gaps
Both platoons and gaps have the geometr
PIplatoon of length k] = q
P[gap of length k] = p(l - p)*"

ACKNOWLEDGMENT
The authors wish to thank Neil Fogo fo
experimental work and Susan Hanson for
and programming aspects of this proje
making available a computer program f
is also acknowledged. The authors have be
Robert Herman of the General Motors R

REFERENCES
1. D. J. Buckley, "Road Traffic Counting D
116 (1967).
2. F. A. Haight, Mathematical Theories of T
York, 1963.
3. G. F. Newell, "Mathematical Models of F
Opns. Res. 3, 176-186 (1955).
4. M. S. Bartlett, "The Spectral Analysis of
Soc, B. 25, 264-296 (1963).
5. D. R. Cox and P. A. W. Lewis, The Statist
Methuen, London, 1966.
6. R. R. Allan, "Extension of the Binomial
tinuous Case," Proc. Aust. Rd. Res. Board 3
7. D. L. Gerlough and A. Schuhl, Poisson
Saugatuck, Conn., 1955.

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms
 A DISCRETE MARKOV MODEL OF VEHICULAR TRAFFIC / 251

8. M. J. Beckmann, C. B. McGuire, and C. B. Winsten, Studies in the Eco

nomies of Transportation, Yale University Press, 1956.
9. J. C. Tanner, "The Sampling of Road Traffic," Appl. Statist. 6,161-170 (1957).
10. A. J. Miller, "An Analysis of Bunching in Rural Two-Lane Traffic," Opns.
Res. 11, 236-247 (1963).
11. K. R. Gabriel and J. Neumann, "A Markov Chain Model for Daily Rainfall
Occurrence at Tel Aviv," Quart. J. Royal Met. Soc. 88, 90-95 (1962).
12. P. A. W. Lewis, "A Computer Program for the Statistical Analysis of Series of
Events," I.B.M. Systems J. 5, 202-225 (1966).
13. J. Durbin, "Some Methods of Constructing Exact Tests," Biometrika 48,
41-55 (1961).
14. D. R. Cox and H. D. Miller, The Theory of Stochastic Processes, Methuen,
London, 1965.
(Received February 1968)

This content downloaded from 14.139.185.114 on Thu, 08 Dec 2016 04:54:28 UTC
All use subject to http://about.jstor.org/terms