1
2
3
Hierarchical Capacity Analysis of Freeways via Nonparametric Bayesian 4
Estimation with Censored Data 5
6
7
8 Eren Erman Ozguven, M.Sc. (Corresponding Author) 9
Graduate Assistant, 10
Department of Civil and Environmental Engineering, 11
Rutgers University 12
623 Bowser Road, Piscataway, NJ 08854 USA, 13
Tel: (732)445-4012 14
Fax: (732)445-0577 15
e-mail: [email protected] 16
17
18
Kaan Ozbay, Ph.D. 19 Professor, 20
Department of Civil and Environmental Engineering, 21
Rutgers University 22
623 Bowser Road, Piscataway, NJ 08854 USA, 23
Tel: (732)445-2792 24
Fax: (732)445-0577 25
e-mail: [email protected] 26
27
28
29 30 31 32 33 34 35 36 37
38
Word count: 5498 + 3 Figures + 5 Tables = 7498 39
Abstract: 195 40
Re-submission Date: November 15, 2010 41
42
43
44 Paper Submitted for Presentation and Publication at the 45
Transportation Research Board’s 90th
Annual Meeting, Washington, D.C., 2011 46
47
TRB 2011 Annual Meeting Paper revised from original submittal.
ABSTRACT 48 49
Traffic capacity of a freeway differs depending on its distinct sections with different 50
spatial characteristics such as the number and width of lanes, existence and type of 51
shoulders and/ or medians, traffic characteristics (such as the number of breakdowns 52
defined using the sudden changes in the speed and density values that occur during the 53
flow phase transition), and population characteristics (rural and urban areas). To account 54
for these spatial differences, this paper investigates the hierarchical estimation of the 55
traffic capacity distribution on a highway using a nonparametric Bayesian approach 56
assuming two prior distributions, namely Dirichlet and Gamma process priors under the 57
minimization of a squared-error loss function. This approach addresses the difficult 58
problem of the censored observations while treating the model parameters as random 59
variables represented by a probability distribution. The methodology is applied on the 60
highway sections with different spatial characteristics. An application of the method for 61
on and off-ramps of a highway is also presented. Finally, the results are discussed 62
hierarchically with presenting a methodology to simulate censored and breakdown 63
observations and to analyze them statistically using a bootstrap approach in order to 64
obtain the capacity distributions for sections without sufficient data. 65
66
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
3
INTRODUCTION 67
Traffic capacity has utmost importance in every phase of the planning, 68
maintenance and operations for various sections of a roadway. The overestimation of the 69
capacity leads to delays and traffic incidents due to the congestion and excessive volumes 70
whereas the underestimation can lead to high costs and unnecessarily large 71
ramp/mainline delays. Thus, there is a need to study the spatial characteristics of the 72
probabilistic distribution of traffic capacity utilizing the traffic state information 73
including free flow and breakdown observations. Although they have not explicitly 74
considered a concept of capacity distribution, several researchers have recognized the 75
necessity of characterizing roadway capacity as a stochastic phenomenon. 76
Lorenz and Elefteriadou defined the traffic capacity by including probabilistic 77
concepts (2), (3). Kim and Elefteriadou (4) performed a simulation analysis to estimate 78
the capacity of 2-lane/2-way highways focusing on the difference from the deterministic 79
capacity values. Persaud et al. (5) studied the probabilistic breakdown phenomenon in 80
freeway traffic and confirmed that breakdown is stochastic in nature. Chung et al. (6), 81
and Hall and Agyemang-Duah (7) worked on the freeway capacity drop and the 82
definition of the capacity whereas Cassidy and Bertini (8), and Smilowitz et al. (9) 83
studied some important traffic features with respect to the bottleneck flows. A 84
deterministic method was proposed by Cassidy and May (10) for estimating the traffic 85
capacity of major freeway weaving sections. Wang et al. (11), and Lertworawnich and 86
Elefteriadou (12) also worked on the capacity of weaving sections. Another approach 87
employing neural network techniques for estimating traffic capacity of weaving segments 88
has been proposed by Awad (13). Yeon et al. (14), and Kuhne et al. (15) worked on the 89
development of a stochastic theory of freeway traffic. Kuhne and Mahnke (16) derived a 90
formulation for obtaining the cumulative distribution for traffic breakdowns defining the 91
breakdown as a car cluster formation process. Kuhne (17) used Weibull curves to match a 92
broad variety of cumulative distribution functions. Uchida and Muniero (18) proposed a 93
traffic capacity and covariance estimation methodology using a macroscopic traffic 94
model based on a generalized car following model. Jun et al. (19) used a simultaneous-95
spline regression model for the joint estimation of traffic variables. Similarly, for the 96
estimation of the key traffic stream parameters, a dual regime model was proposed by 97
Yao et al. (20) whereas Banks (21), (22), (23), (24) conducted several important studies 98
to develop a new approach to bottleneck capacity analysis. Polus and Pollatschek (25) 99
explored the stochastic nature of freeway capacity by fitting speed-flow diagrams. Laval 100
(26) came up with approximate formulas for highway capacity using the presence of slow 101
vehicles. 102
However, none of these studies focus on the use of statistically robust 103
methodologies to estimate the probabilistic distribution of the freeway capacity spatially. 104
Brilon et al. (27), (28) used a practical estimation method based on an analogy to the 105
statistics of lifetime analysis to obtain the distribution of freeway capacity. They used the 106
Kaplan-Meier estimator to obtain the capacity distribution function, and estimated the 107
parameters of this distribution with various functions, where Weibull function appeared 108
to be the best fit. This lifetime method had been first proposed by van Toorenburg (29) 109
and discussed by Minderhoud et al. (30). Regler (31), applied the capacity analysis in 110
(28) to other freeway sections. Recently, Geistefeldt and Brilon (1) and Geistefeldt (32) 111
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
4
made a comparative assessment between the direct estimation of breakdown probabilities 112
and estimation of capacity distribution functions via Kaplan-Meier method. They 113
concluded that Kaplan-Meier method performs better than the direct breakdown 114
probability estimation techniques. Ozbay and Ozguven (33), studied the effects of 115
spatial/temporal differences of freeway sections on the capacity distribution using a non-116
informative Bayesian prior distribution called Jeffrey's prior. Ozguven and Ozbay (34), 117
improving the method given in (27), introduced a nonparametric Bayesian estimation 118
under an informative Dirichlet process prior, for the first time, to robustly estimate the 119
traffic capacity model parameters in the presence of insufficient and unreliable data. 120
In this article, we make the following contributions: 121
Improving the Bayesian estimation performed by Ozguven and Ozbay (34), we 122
propose a hierarchical nonparametric Bayesian approach to estimate the traffic 123
capacity distribution where the prior knowledge about the traffic parameters specific 124
to a freeway section is represented by a probabilistic function. 125
We use two different probability distributions, namely Dirichlet and Gamma 126
processes as the priors to compare the traffic capacity distributions for different 127
sections of the same highway, and to test the location specific differences among the 128
estimated distributions. 129
We apply this estimation methodology to the ramp sections using the occupancy 130
data which can be used to enhance freeway control strategies such as ramp metering. 131
With the proposed model, a freeway can be studied hierarchically with respect to 132
the following factors: 133
Number of lanes, 134
Population affects (rural and urban areas), 135
Traffic features (number of breakdowns, free flows and flow phase 136
transition). 137
HIERARCHICAL BAYESIAN ESTIMATION WITH CENSORED DATA 138
The use of Bayesian estimation to obtain the traffic capacity is useful when the 139
following conditions are met: 140
There is substantial amount of censored data in the traffic flow observations. 141
There is insufficient data to study the behavior of the traffic capacity accurately 142
and efficiently. 143
Traffic observations are not 100% reliable. 144
In the context of the probabilistic traffic capacity modeling mentioned in previous 145
studies, these conditions are present. Therefore, there is definitely a need for using 146
hierarchical Bayesian estimators. 147
We first consider the following model. Let 1 2( , ,..., )nq q q q be the traffic flow 148
data and 1 2( , ,..., )nV V V V be the speed data of n observations obtained from the sensors 149
on a freeway section through a specific time period. Data is censored from the right by a 150
threshold speed value of Y where q consists of breakdown and censored flows. 151
Censored flows are the ones where the speed values are higher than the threshold speed 152
so that they do not cause breakdown. Therefore, the data is in the form of: 153
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
5
, 1,...,
1 if
0 if
j
j
j
j
q j n
V Y
V Y
(1) 154
If 1j , there is a traffic breakdown at volume jq , and, if 0j , traffic is 155
fluent above the threshold speed without any breakdown (censored). Therefore, the 156
important point is to make use of the traffic observations, including the censored 157
volumes, to calculate the probability distribution of the capacity. With this information, 158
we obtain the capacity distribution function ( )F q as: 159
( ) 1 ( ) ( )F q S q P C q (2) 160
Here, C represents the section capacity, and ( ),S q the survival function. The 161
Kaplan-Meier estimation approach to find this capacity distribution was presented in 162
(27). The difference of Bayesian statistics from this approach is that it incorporates the 163
prior knowledge along with a given set of current observations to make statistical 164
inferences. The prior information could come from operational/observational data, from 165
previous experiments or engineering knowledge. The pioneering studies on the Bayesian 166
estimation with right-censored data were conducted by Susarla and Ryzin (35), and 167
Ferguson and Phadia (36). The idea is to obtain the nonparametric estimation of the 168
capacity distribution function using a Bayesian approach with right-censored 169
observations. As a base decision rule to obtain an estimate of the traffic capacity function, 170
a loss function is specified as 171
2
0
( , ) ( ( ) ( )) ( )L F F F q F q dw q
(3) 172
where w is a nonnegative weight function. This loss function simply presents the 173
weighted integrated difference between the actual and estimated values of the capacity 174
function. 175
Kaplan-Meier Estimation (Product Limit Method) 176
The standard estimator of the survival function is the Kaplan-Meier estimator 177
(Product-Limit Estimator) proposed by Kaplan and Meier (37), and applied into traffic 178
engineering by Brilon et al. (27) to obtain the distribution of freeway capacity. Using this 179
approach, traffic capacity distribution function, ( )KF q is obtained as: 180
:
( ) 1 ( ) 1j
j j
K K
j q q j
nF q S q
n
(4) 181
where 182
( )KF q : estimated Kaplan-Meier capacity distribution function, 183
( )KS q : estimated Kaplan-Meier survival function, 184
jn : number of intervals with a flow rate jq , 185
jq : traffic volume in interval j , 186
j : number of breakdowns at a volume of jq . 187
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
6
This product is calculated over all time intervals j with a traffic volume jq q 188
that were followed by a traffic breakdown. With a comparative assessment, Geistefeldt 189
and Brilon (1) concluded that Kaplan-Meier method performs better than the direct 190
breakdown probability estimation techniques. 191
Hierarchical Bayesian Approach 192
In this paper, a nonparametric hierarchical Bayesian estimation of the traffic 193
capacity analysis is given under the squared-error loss notion with stochastic process 194
priors. We consider a nonparametric approach in which 0 ( )F q , the prior guess on the 195
estimated capacity function, is drawn from a stochastic process ( , )f H based on the 196
observations ( , )q . Here, 0 and H are the baseline probability measures for 197
( , )f H . With this information, the posterior distribution of the capacity distribution 198
function, ( )F q is derived. Therefore, the hierarchical stochastic process becomes (see 199
Teh et al. (38) for details): 200
0
0 0
( ) , ( , )
( ) , ( ) ( , ( ))
F q H f H
F q F q f F q
(5) 201
where is a parameter of the prior distribution 0 ( )F q . The selection of the prior guess 202
has utmost importance. This prior guess based on ( , )f H can be obtained from previous 203
empirical work, data, or researcher's subjective beliefs. In the proceeding sections, we try 204
two different priors, namely Dirichlet and Gamma process priors, to obtain the posterior 205
distribution of the traffic capacity function. 206
Dirichlet process prior 207
Dirichlet distribution is the first prior distribution selected to lead to closed-form 208
estimates for the survival function using the loss function given in Equation (3). This 209
prior is chosen because it is a conjugate prior (the one which produces a posterior 210
distribution of the same type as the prior) for the survival function. This makes the 211
computation of the posterior particularly simple. 212
When we use a Dirichlet process (DP) in Equation (5), the general hierarchical 213
model becomes as follows: 214
0
0 0
( ) (0, ) ( (0, ))
( ) , ( ) ( , ( ))D
F q DP
F q F q DP F q
(6) 215
To assign a prior distribution to the capacity function, the parameter function is 216
taken to be of the form 0( , ) (1 ( ))q F q where 0( ) 1 (0, )F q is the prior 217
guess at the traffic capacity function and is a kind of measure that indicates how much 218
weight to put on the prior guess. Here, we take the prior guess at the survival function 219
0 0( ) 1 ( )S q F q with a Dirichlet prior given as (0, ) qe and our problem becomes: 220
0 ( ) 1 (0, ) 1
( ) , ( , ) ( , ( , ))
q
D
F q e
F q q DP q
(7) 221
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
7
As observed, there is a precise relationship between the prior and posterior 222
distributions; therefore, we obtain the posterior distribution as: 223
0( , ) (1 ( )) (0, ) qq F q e (8) 224
where is heuristically calculated using the Kaplan-Meier data of breakdown intervals. 225
To find the Bayesian estimator with Dirichlet prior, the conditional pth
moment of 226
the breakdown probability ( ) 1 ( )D DF q S q given ( , )q is calculated as follows, where 227
( , )q is distributed as a Dirichlet process on (0, )R : 228
1
0
( , )(1 ( )) ( , ) ( )
(0, )
pqp
D
s
q s nE F q q j
s n
(9)
229
which is adapted from (35) where 230
( , ) :
( , )( )
un : 1
j j
j
j j j j
j
q nwhile q censored
q nj
while q censored
231
jn : number of intervals with a flow rate greater than or equal to jq , 232
n : total number of observations, 233
qn : total number of observations with a flow rate greater than q , 234
j : number of censored observations at a volume of jq . 235
jq : traffic volume in interval j , 236
p , s : nonnegative integers. 237
Using Equation (9), the Bayesian estimator of the survival function with the 238
Dirichlet prior is calculated taking 1p . That is, we are focusing on the conditional 239
mean of 1 ( )DF q given ( , )q . Then, our Bayesian estimator becomes: 240
( , )( ) 1 ( ) 1 ( )
(0, )
q
D D
q nF q S q j
n
(10) 241
where 242
( , ) :
( , )( )
un : 1
j j
j
j j j j
j
q nwhile q censored
q nj
while q censored
243
( )BS q : estimated Bayesian survival function. 244
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
8
Gamma process prior 245
To compare the Dirichlet results with another distribution, we choose the Gamma 246
prior, namely ( )GF q . With the Gamma prior having a continuous shape parameter ( )q 247
and reciprocal scale parameter , the posterior expectation is calculated for the traffic 248
capacity as follows (see (36) for the original formulation): 249
( )( )
1
1 1
1 ( 1, )(1 ( )) ( , )
1 ( , )1
jq
qqn n
j jq G j j j
G
jq G j j jj j
n nn nE F q q
n nn n
250
(11)
251
where 252
jn : number of intervals with a flow rate jq , 253
n : total number of observations, 254
qn : total number of observations with a flow rate greater than q , 255
1 if
0 if
j
j
j
V Y
V Y
256
j : number of censored observations at a volume of jq . 257
jq : traffic volume in interval j , 258
: scale parameter of the Gamma prior, 259
( )q : shape parameter of the Gamma prior, 260
1
0
1 1( , ) 1 log( )
i
G
i
i
i i
. 261
Here, the prior Gamma guess at the capacity function is selected as: 262
( )
0 0( ) 1 ( ) 11
q
F q S q
(12) 263
Then, for all q , we have ( )q for fixed as: 264
0ln(1 ( ))( )
ln1
F qq
(13) 265
266
Finally, the Bayesian capacity function estimator with the Gamma prior is 267
calculated as in the Dirichlet prior case, and our Bayesian estimator becomes: 268
( )( )
1
1 1
1 ( 1, )( ) 1 ( ) 1
1 ( , )1
jq
qqn n
j jq G j j j
G G
jq G j j jj j
n nn nF q S q
n nn n
269
(14) 270
where ( )q is given as in Equation (13) and ( )G as in Equation (11). 271
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
9
The Bayesian estimator, using the Dirichlet or Gamma process priors, estimates 272
all possible probability values between the breakdown and censored observations 273
(achieving continuity) and therefore smoothes the posterior distribution curve at all 274
discontinuities. Suzarla and Ryzin (39) state that "the Bayesian estimator is a better 275
admissible estimator smoothing the resulting posterior on the discrete survival data using 276
the expectation and variances of the cumulative distribution given in Equation (9) under 277
the loss function used and under the weak convergence topology". Survival models are 278
based on data that measure time to some event such as death, transition or failure. In our 279
case, this time period is replaced by the traffic volume and the discrete failure event 280
becomes the discrete traffic breakdown event. Therefore, assuming the breakdown as a 281
failure event, the statistical methods developed for the survival data analysis can be 282
successfully used to estimate the traffic capacity, which is the analog of the lifetime in 283
this context. 284
COMPARATIVE MODEL APPLICATION 285
Basic Assumptions 286
The concept of determining how to define the breakdown is very crucial in terms 287
of the accuracy and efficiency of the proposed methodology. There are several 288
approaches available to define the freeway breakdown. One approach is to determine the 289
breakdown value by using the speed difference between consecutive intervals whereas 290
another approach is including the average space mean speed to detect the breakdown 291
point. Recently, Geistefeldt and Brilon (1) used a more sophisticated set of criteria to 292
determine the breakdown value. In this study, the flow values are classified, using a 5-293
minute interval denoted as i , and shown in Table 1 where Y is a threshold speed, and 294
is a small speed value selected for comparison to determine the breakdown. This 295
approach is logical as the traffic breakdown is usually followed by a significant amount 296
of reduction in speed, however every interval followed by a speed lower than the 297
threshold value cannot be considered as a breakdown interval (B). If each congested 298
volume was considered as a B-value, we would be considering all the breakdown flows 299
again and again until the end of the analysis after the failure of a specific road segment. 300
Hence, the intervals that causes breakdown with satisfying the conditions in Table 1 is 301
considered as a B-value. 302
The traffic congestion usually leads to a breakdown, which is usually determined 303
by a sudden decrease in the speed level accompanied with a simultaneous steep increase 304
of the traffic density level that occurs during the phase transition from free-flow to 305
breakdown flow. When the traffic breakdown ends, the speed and density both return to 306
the normal level of the uncongested conditions. Therefore, while considering the land 307
usage for the proposed spatial analysis in this study, the phase transition is also 308
considered during the determination of the breakdown points used in Table 1. 309
310
311
312
313
314 315
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
10
TABLE 1 Breakdown Interval Definitions (Based on (1)
) 316
Interval Definition
B
Traffic is fluent in time interval i , but the observed volume causes a
breakdown, satisfying the following conditions:
1iv Y ,
iv Y ,
1iv Y ,
2iv Y ,
1 1 2
1 1( ) ( )
2 2i i i iv v v v .
C
Traffic is fluent in the intervals i and 1i . This interval i contains a
censored value. Its information is that the actual capacity in interval i is
greater than the observed volume iq . C-values represent the right-censored
data in the analysis. That is, C-values are accompanied by speed values
higher than the threshold speed and they do not cause breakdown.
D1
Traffic is congested in interval i , i.e. the average speed is below the
threshold value. As this interval i provides no information about the
capacity, it is disregarded. These are mainly those intervals representing
congested flow conditions, which do not contain any information about the
traffic capacity.
D2
D2-values are similar to B-values; however they are due to a breakdown in
a downstream cross-section. That is, the reason that the speed goes below
the threshold value is not the interval itself, but rather downstream
congestion. Traffic is fluent in interval i , but the observed volume causes a
breakdown. However, unlike classification B, traffic is congested at a
downstream cross section during interval i or 1i , this case, the
breakdown at the observation point is supposed to be due to a tailback from
downstream. As this interval i does not contain any information for the
capacity assessment at the observation point, it is disregarded.
D3
Traffic data may include low traffic flow values having a wide range of
corresponding speed values. As they have nothing to do with the
breakdown, they are disregarded within a certain threshold for the data. 317
Data 318
Data obtained from different sections of the I-5 Highway, California, USA from 319
the PEMS database (40) is used to test the location specific differences among the 320
estimated distributions. The information about the sections studied are given in Table 2. 321
322
323
324
325
326
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
11
TABLE 2 I-5 Highway Section Information 327
Vehicle
Detection
System (VDS)
Number
Type Number of
Lanes Location
Maximum
Observed
Volume per
lane (PEMS(40)
)
1015010 Mainline 2 Rural (Stanislaus
County, CA) 1524 veh/h
1108613 Mainline 4
Urban
(Washington St,
San Diego, CA)
1914 veh/h
1205157 Mainline 6 Urban (Orange
County, CA) 1732 veh/h
1214241 On Ramp 2 Urban (Orange
County, CA) -
1211850 Off Ramp 1 Urban (Orange
County, CA) -
328
The important point of the study while using the data is to obtain plausible 329
cumulative distribution functions for different sections of the freeway (which results in 330
obtaining data from several single loop detectors on different lanes of the freeway 331
sections) so that this can be used to make it possible for a traffic engineer/planner to 332
choose the proper pairs of the breakdown probability and traffic capacity for the whole 333
freeway. After analyzing one-year (2009) traffic data obtained for these highway 334
sections, threshold speeds for the mainline sections are selected as 70 m/h , whereas it 335
was taken as 40 m/h for the ramp sections, and is selected 5 m/h for all sections. 336
We have also checked the validity of the breakdown speed values from 337
consecutive detectors. For instance, as observed from Table 2, Vehicle Detection System 338
(VDS) No. 1108613 is one of the detection systems used in the analysis. However, one of 339
the consecutive systems namely, VDS-1108603 is also tested using the model to 340
determine if the breakdowns and the corresponding traffic flow per lane values are 341
reasonably similar. This is simply a way of validating the data. 342
Example Calculation 343
Firstly, we work on an example to illustrate the concept of the Bayesian 344
estimation. Kaplan-Meier estimator is indeed a limit of the Bayesian estimator given in 345
Equation (6) when ( ) 0R . This reduction and better performance of Bayesian 346
estimation using censored data is clearly seen on an example with the following data (8 347
observations obtained from VDS-1015010): 348
: 40,70,90,120 veh/5 min
:50,60,100,130 veh/5 min
Breakdown flows
Censored flows 349
Given the data and selecting the prior guess of the survival function as 350
0 0( ) 1 ( ) qS q F q e , the Bayesian estimators and Kaplan-Meier estimates are 351
calculated in Table 3. Based on the prior guess of the survival function, and taking 1 , 352
we obtain ( ) 1.443q q from Equation (13). 353
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
12
TABLE 3 Bayesian and Kaplan-Meier Estimates of ( )S q for the Example 354
Flow
(veh/5
min) in
the
interval
Bayesian
Estimate with Dirichlet prior
ˆ ( )DS q
Bayesian
Estimate with Gamma prior
ˆ ( )GS q
Kaplan-
Meier
Estimate
ˆ ( )KS q
[0,40) 8
8
qe
1.449
10
q
1.0
[40,50) 7
8
qe
1.44 57.7210
ln( )8 81 9* *
99 80ln( )
8
q
7/8
[50,60)
50
50
6 7*
8 6
qe e
e
1.44 57.72 72.1510
ln( )7 81 649* * *
98 80 63ln( )
8
q
7/8
[60,70)
50 60
50 60
3 7 6* *
8 6 5
qe e e
e e
1.44 86.586 49
*......*7 48
q
7/8
[70,90)
50 60
50 60
3 7 6* *
8 6 5
qe e e
e e
1.44 1017
ln( )5 36 6*......*
66 35ln( )
5
q
7/10
[90,100)
50 60
50 60
3 7 6* *
8 6 5
qe e e
e e
1.44 129.876
ln( )4 25 5*......*
55 24ln( )
4
q
21/40
[100,120)
50 60
50 60
100
100
2 7 6* *
8 6 5
3*
2
qe e e
e e
e
e
1.44 144.3
3 16*......*
4 15
q
21/40
[120,130)
50 60
50 60
100
100
2 7 6* *
8 6 5
3*
2
qe e e
e e
e
e
1.44 173.16
4ln( )
2 9 3*......*33 8
ln( )2
q
21/80
[130, )
50 60
50 60
100 130
100 130
7 6* *
8 6 5
3 1* *
2
qe e e
e e
e e
e e
187.59
4*......*
3
qe
Not
defined
355
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
13
As ( ) 0Rq
R e
, the ratios coming from the censored data naturally 356
vanish as the values common to both products cancel out reducing the Bayesian estimator 357
with Dirichlet prior to the Kaplan-Meier. However, the Gamma prior estimate does not 358
converge to the Kaplan-Meier. It converges to an estimate which gives more weight to 359
the right tails than the Kaplan-Meier estimate. This takes the Bayesian estimator with 360
Gamma prior away from the other two estimators. 361
Comparative Spatial Results 362
Distributions for the Dirichlet and Gamma priors, namely (0, ) and
( )
1
q
363
are calculated based on the following heuristic argument given in (35) for each freeway 364
section separately. The idea is to force the estimators to satisfy the following equation: 365
,0.5M
Ke quantile P (15) 366
where : ,0.5KM quantile P for the estimated Kaplan-Meier breakdown probability 367
vector KP . Therefore, M represents the 0.5 quartile of the Kaplan-Meier estimated flow 368
vector. Using this method, and by doubly using the observations, is obtained for each 369
section. This approach is found to be useful in (35) as other values of tend to pull the 370
Bayesian estimator away from the Product-Limit estimator. 371
The selected parameters for the Gamma and Dirichlet distributions are given in 372
Table 4. The parameters for Dirichlet prior and for Gamma prior define the prior 373
mass on the traffic volume data. For instance, when the data sample size is n , the prior 374
mass on (0, ) becomes n
. The extreme case is to let in which case the 375
Bayesian estimator reduces to qe . values are optimally obtained for each section 376
iteratively tending not to pull the Bayesian estimator from the Kaplan-Meier curve. On 377
the other hand, the intensity parameter is selected as 1 according to the argument given 378
in (36) and by choosing that value, we correspond reasonably well to the prior sample 379
size for the Gamma process given in this study. With this information, ( )q is calculated 380
with Equation (13). 381
382
383
384
385
386
387
388
389
390
391
392
393
394
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
14
TABLE 4 Distribution Parameters for the Highway Sections Studied 395
Section ( )q Sample Size
( )q
VDS-1205157 (6 Lanes) 0.1071 0.1071qe 0.0050
0.1071ln( )
ln1
qe
VDS-1108613 (4 Lanes) 0.0498 0.0498qe 0.0040
0.0498ln( )
ln1
qe
VDS-1015010 (2 Lanes) 0.0126 0.0126qe 0.0045
0.0126ln( )
ln1
qe
VDS-1211850 (1 Lane) 0.0344 0.0344qe 0.0035
0.0344ln( )
ln1
qe
VDS-1214241 (2 Lanes) 0.0536 0.0536qe 0.0048
0.0536ln( )
ln1
qe
396
Figure 1 represents the estimated capacity distribution curves for the 2-lane, 4-397
lane and 6-lane sections of I-5 Highway and the spatial comparative plot for the 398
estimators obtained using these parameters. 399
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
15
400 (a) 401
402 (b) 403
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
16
404 (c) 405
406 (d) 407
FIGURE 1 Capacity estimators of I-5 highway (FIGURE 1a) for section 1, 408
(FIGURE 1b) for section 2, (FIGURE 1c) for section 3, and (FIGURE 1d) capacity 409
comparison with Bayesian estimation for all sections 410
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
17
The maximum value of the Kaplan-Meier curve only reaches 1 if the maximum 411
observed volume q is followed by a breakdown. This is encountered in (27), where 412
despite the used large sample, no complete distribution function could be obtained since 413
the highest flow values observed were not followed by breakdown. Bayesian estimators 414
solve this problem and give more complete curves for the estimated capacity 415
distributions. To sum up, the lower and upper tails of the capacity distribution cannot be 416
estimated by non-Bayesian estimation techniques. 417
Moreover, Gamma prior does not work as well as the Dirichlet prior. The 418
Dirichlet prior estimate is closer to the Kaplan-Meier estimate than the Gamma prior 419
estimate. This is because the estimator with Dirichlet prior assigns more mass on the 420
breakdown flows and less mass between these flows and at the tails than Gamma prior 421
estimate. Unlike for the Dirichlet process where the relative change due to a breakdown 422
in the estimator occurs only at the observation itself, for the Gamma prior, some of the 423
information obtained from a breakdown is used to change the relative weight of the 424
points of the estimator to the left of it. Therefore, based on the observations obtained 425
from empirical data, Dirichlet distribution is chosen as a reasonable prior for our model 426
that represents the distribution of traffic capacity. 427
For the Kaplan-Meier estimator, the breakdown probability is 0 up-to the flow 428
with first breakdown, which is not realistic for every case. For instance, consider the 429
section with 4-lanes (VDS-1108613) of I-5 Highway. With a flow value of 1959 veh/h, 430
the Kaplan-Meier curve has a breakdown probability of 0.33 roughly whereas the 431
breakdown probability is 0 for 1515 veh/h. This means that the traffic is almost free 432
flowing (no chance of breakdown) at that volume which is merely 444 veh/h less than 433
1959 veh/h. This unrealistic situation is addressed by the Bayesian estimator calculating 434
the breakdown probabilities for all traffic flow values possible including the relatively 435
lower flows for which breakdowns are not actually observed. Therefore, Bayesian 436
estimation becomes more effective while defining the capacity profile of a freeway 437
section. 438
Bayesian estimators are smoother than the Kaplan-Meier estimator in the sense 439
that the jumps at censored observations are not as large for the Bayesian estimators. 440
Actually, both the conventional and Bayesian estimate are discontinuous at uncensored 441
observations. However, Bayesian approach smoothes the estimator at censored 442
observations by achieving continuity. This smoothness depends on the value of , and 443
the percentage of censored observations which differs from section to section. For 444
example, for the rural section (VDS-1015010), the breakdowns are very limited, and at 445
the same time, the censored observations for higher traffic volumes are not present. This 446
makes it even difficult for the Bayesian estimation to obtain the upper part of the curve. 447
Therefore, possible trajectories (trend lines) for the sections without enough data are 448
plotted based on the shape of the section with VDS-1205157 which has the satisfactory 449
data. A procedure to obtain a trend line mathematically is given in the proceeding 450
sections. 451
Application to Ramps 452
PEMS ramp data (40) do not include the density and volume observations. It only 453
has the number of vehicles passing through the detector, and their occupancies. 454
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
18
Therefore, we apply the following formula to calculate the speed values from the single-455
loop detector data (41): 456
ˆ( )DL LQV
O
(16) 457
where 458
Q : number of vehicles per 5 minutes, 459
O : occupancy, percentage of the time the loop is occupied by vehicles 460
during the 5 minutes interval, 461
L̂ : average vehicle length, 462
DL : length of detector, 463
: conversion constant. 464
Choosing the average vehicle length as 16 ft, and the length of detectors obtained 465
from the PEMS data, we obtain the speed observations for the ramp sections where Y is 466
selected as 40 m/h , and as 5 m/h . Despite the large data size, it is difficult to come up 467
with a reasonable amount of traffic observations (both breakdowns and censored 468
observations) for the off-ramp section using Table 1. The breakdown values are very 469
limited, and the largest flow value is followed by a breakdown. Hence, Kaplan-Meier and 470
Bayesian estimator curves are almost the same. For VDS-1214241 (on-ramp with 2-471
lanes), we have better data to obtain the capacity distribution. This is possibly because the 472
section with 2-lanes is more suitable to apply a breakdown analysis suggested in this 473
paper. Actually, the section can be viewed as a short mainline freeway section. Moreover, 474
the comparison figure indicates that the capacity of the urban on-ramp is significantly 475
higher than the urban off-ramp section. 476
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
19
477
478 (a) 479
480 (b) 481
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
20
482 (c) 483
FIGURE 2 Capacity estimators of two ramp sections of I-5 highway (FIGURE 2a) 484
for section 1, (FIGURE 2b) for section 2, and (FIGURE 2c) capacity comparison 485
with Bayesian estimation for all sections 486 487
Ramp metering strategies that heavily depend on the accurate estimation of the 488
section capacities can be significantly improved using this methodology. For example, 489
for the VDS-1214241 section, rather than having just one probability value for the 490
capacity, a capacity range from 150 to 250 veh/h/lane can be used as (Figure 2): 491
Capacity interval: 150 185 Breakdown probability: 0.25 492
Capacity interval: 185 210 Breakdown probability: 0.50 493
Capacity interval: 210 250 Breakdown probability: 0.90 494
These intervals indicate that the breakdown probability can occur for any traffic 495
flow value between the boundaries of these intervals. For instance, knowing that VDS-496
1214241 section has a 0.50 probability of breakdown for 185 to 210 veh/h/lane, it is 497
possible to adjust the traffic signaling system at the end of the ramp accounting for the 498
higher flows. 499
Discussion of the Hierarchical Modeling Approach 500
The choice of the traffic capacity depends on the engineer’s judgment/purposes at 501
the specific period s/he is working in. In general, the main idea is to choose a strategy 502
that ensures the maximum flow with minimum breakdown probability. Moreover, the 503
breakdown probability initially selected at the design stage may change due to 504
operational characteristics, external conditions, or even a new construction. Using the 505
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
21
presented methodology, the operating agency has the opportunity to revise the breakdown 506
probability. 507
Spatially, the steepness of the curve is related to the quality and quantity of the 508
data, which changes the parameters of the posterior distributions. For instance, 509
parameter directly depends on the number of breakdown and censored observations. As a 510
rule of thumb, the larger it is, a better data set (better representative information) we have 511
for the roadway section. The parameter for Dirichlet prior, on the other hand, defines 512
the prior mass on the traffic volume data. When the sample size is n , the prior mass on 513
(0, ) becomes n
. The larger it is, the smoother the Bayesian estimator becomes when 514
compared to the Kaplan-Meier estimator. 515
To calculate the Kaplan-Meier estimator, we do not have to know the actual 516
censored observations, but the number of censored observations between two uncensored 517
traffic volumes is sufficient. The Bayesian estimators, however, are calculated with using 518
all the data including both uncensored and censored observations. The fact that Bayesian 519
estimates use all the data makes it preferable to the other estimates. In this sense, the 520
Bayesian estimators serve as a function of the sufficient statistic by fully utilizing the 521
data. 522
The spatial results presented indicate the need for a hierarchical sampling 523
approach to obtain the freeway capacity curves incorporating the stochastic nature of the 524
traffic volume and the section characteristics. For instance, for the rural section of I-5 525
highway (VDS-1015010), it is difficult to obtain a full curve because data obtained do 526
not include reasonable amount of breakdown and even censored observations (Figure 1). 527
On the other hand, for the urban section of I-5 highway (VDS-1205157), we have a better 528
capacity distribution estimate where substantial amount of data including breakdowns are 529
present. Although these two sections differ in terms of driver population characteristics 530
and number of lanes, the lanes were actually built with similar design principles, and 531
therefore the actual capacity distribution curves of the roadways should be similar in 532
shape. Therefore, given the sufficient amount of data, it can be possible to obtain a 533
capacity estimator curve for VDS-1015010 by changing the parameters of the capacity 534
estimating distribution. With this idea, data obtained for the urban section (with sufficient 535
amount of breakdown and censored observations) is used as the base to simulate the prior 536
distribution of the hierarchical model presented. This is also a way of assessing the 537
robustness of the estimation results by analyzing the subsets of the traffic data samples. 538
Geistefeldt and Brilon (1) chose half of their data set as the subset sample, and obtained 539
the estimator curves calculating the parameters of a Weibull distribution. Similarly, we 540
focus on the data obtained from VDS-1205171, and we take a sample of 3 -months from 541
the data set to obtain reliable parameters for the Dirichlet estimator. 542
To recall, the data should be in the form of: 543
, 1,...,
1 if
0 if
j
j
j
j
q j n
V Y
V Y
544
If 1j , there is a traffic breakdown at volume jq , and, if 0j , traffic is fluent above 545
the threshold speed without any breakdown (censored). From Equation (10), the Bayesian 546
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
22
estimate of the capacity curve for any breakdown Bq value with the Dirichlet prior, 547
ˆ( ( ))BF q is given as follows: 548
( , )ˆ ( ) 1
(0, )
q
B
q nF q
n
(17) 549
whereas the estimate for the censored observations ( ˆ ( )CF q ) for any censored value Cq is 550
as follows: 551
:
( , ) ( , )ˆ ( ) 1
(0, ) ( , )j
q j j
C
j q censored j j j
q n q nF q
n q n
(18) 552
We obtain these breakdown and censored observations, namely Bq and
Cq , where 0 553
or 1 as Bq q or Cq q , respectively. Then, we estimate the parameter values , to 554
obtain ˆ ( )CF q and ˆ ( )BF q based on the observations ( , )q of the VDS-1205157 section 555
where ( , )q is the empirical distribution function assigning the prior mass on each 556
observed value of traffic volume in (0, ) as n
. To perform that, we employ the well-557
known bootstrap method (Efron (42)) and assign a measure of accuracy to the estimated 558
parameters. The bootstrap methodology used for this model can be described as follows: 559
560
We randomly obtain a 3-months data sample of size l from the VDS-1205157. In 561
this data set, ˆ ( )BF q puts mass only at those observations Bq q where 1j , 562
and ˆ ( )CF q puts mass only at those observations Cq q where 0j . 563
This gives a bootstrap empirical distribution function *( , )q , the empirical 564
sample distribution , 1,...,jq j l , and the corresponding bootstrap value 565
* *ˆ ( )F F q . 566
These steps are repeated independently for a large number of iterations, say N , 567
obtaining bootstrap values *1 *2 *ˆ ˆ ˆ, ,..., NF F F . 568
The value of the measure of accuracy ˆˆ ( )BOOT F is approximately calculated 569
using the sample standard deviation of the *F̂ values: 570
2
*
1* 2
1
ˆ( )
ˆ( )
ˆ1
Nj
Njj
j
BOOT
F
FN
N
(19) 571
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
23
With this approach, we obtain the bootstrap distribution percentiles *F̂ , and 572
standard deviation estimates based on those percentiles. We select N as 1000, and use 573
the Monte Carlo simulation procedure to obtain values in Table 5. The results indicate 574
that the sampling estimator performs sufficient enough in terms of representing the 575
original Bayesian capacity distribution using the VDS-1205157 data. However, it is 576
important to note that the reduced data samples still do contain a considerable amount of 577
at least 95 breakdown observations. That is actually why this section is selected as an 578
input for the analysis. For the data samples including only a few traffic breakdowns, it 579
will be harder to maintain a reasonable sampling bootstrap distribution. 580
581
TABLE 5 Bootstrap Distribution Results 582
q 278q 548q 818q 1097q 1371q 1643q 1918q 2250q
2465q
260 514 769 1042 1297 1550 1803 2125 2340
ˆ ( )F q 0.053 0.056 0.059 0.062 0.071 0.097 0.218 0.449 0.995
ˆBOOT 0.088 0.070 0.065 0.058 0.051 0.046 0.042 0.037 0.031
Percentiles of the bootstrap distribution of *ˆ ( )F q
10% 0.048 0.052 0.056 0.058 0.068 0.082 0.153 0.367 0.876
25% 0.049 0.054 0.058 0.060 0.070 0.089 0.193 0.402 0.914
50% 0.052 0.056 0.060 0.062 0.072 0.095 0.224 0.443 0.995
75% 0.055 0.057 0.062 0.065 0.075 0.102 0.301 0.496 0.996
90% 0.057 0.061 0.064 0.068 0.077 0.107 0.367 0.537 0.999
583
Now, we change the parameters , of the sampling distribution to obtain a 584
reasonable capacity distribution curve for VDS-1015010. The selected values for the 585
parameters , are given in Table 3 for VDS-1015010 roughly representing the 586
sectional and traffic characteristics of the section. As observed in Table 5, the simulated 587
distribution has breakdown probability values similar to those of the input data set. The 588
issue is to which extent the sampled capacity distribution can estimate the capacities for 589
VDS-1015010. Figure 3 shows the actual Bayesian estimator and Kaplan-Meier curve 590
obtained for VDS-1205157, the actual Kaplan-Meier curve for VDS-1015010, and its 591
estimated bootstrap curve obtained via sampling from the input data of VDS-1205157. 592
The result shows that the new sampled curve gives a reasonably good fit for VDS-593
1015010 curve. This can be interpreted as if we had sufficient data for that section to 594
create a capacity distribution curve. 595
596
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
24
597 FIGURE 3 Capacity estimator for VDS-1015010 via Sampling 598
CONCLUSIONS AND FUTURE RESEARCH 599
This paper proposes a hierarchical estimation methodology to determine roadway 600
traffic capacity distribution using nonparametric Bayesian estimation techniques based on 601
two different priors, namely Dirichlet and Gamma process priors. The distribution of 602
traffic capacity in the presence of censored observations is studied incorporating the 603
spatial features of the roadway sections. Then, Kaplan-Meier estimator is compared with 604
the Bayesian estimators. The analysis is based on five-minute interval observations from 605
I-5 Highway in California, USA. The results indicate that, especially for urban roadways 606
and ramp sections where breakdowns are most likely to occur, the use of a stochastic 607
approach is necessary and crucial. 608
It is mostly impossible to obtain the actual capacity distribution as the flow 609
observations have right-censored data points. However, several non-parametric 610
estimation methods that can handle data with censored observations are available in the 611
literature. Kaplan-Meier method is the most well-known estimation technique to obtain 612
the capacity distribution (Geistefeldt and Brilon (1)). In the presence of sufficient amount 613
of data, these techniques can work efficiently. The comparative results presented in this 614
paper suggest that, for a section with sufficient amount of breakdown values, and without 615
too many censored volume values, it is possible to use the Kaplan-Meier estimator. 616
However, in cases where there are substantial amount of censored traffic volumes in the 617
data, it is necessary to use a hierarchical Bayesian estimation approach to obtain a better 618
capacity estimation curve. As shown before, Kaplan-Meier estimator is indeed a limit of 619
the Bayesian estimator with the Dirichlet prior. Thus, we can state that the estimators 620
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
25
complement each other, and the traffic engineer should be careful to decide which one to 621
use for the specific case on hand. Moreover, the Kaplan-Meier estimator is not defined 622
for all possible values of traffic flow values ranging from zero to infinity. Since a 623
complete curve cannot be obtained, parametric distribution fitting techniques have to be 624
applied on the Kaplan-Meier estimator. On the other hand, the proposed non-parametric 625
Bayesian estimators serve as a function of the sufficient statistic by fully utilizing the 626
data. 627
Moreover, it is found that Gamma prior performs worse than Dirichlet prior. This 628
is basically because Dirichlet prior estimator assigns more weight on the breakdown 629
flows and less weight between these flows and at the tails than the one with Gamma 630
prior. Unlike for the Dirichlet process where the relative change due to a breakdown in 631
the estimator occurs only at the observation itself, for the Gamma prior, some of the 632
information obtained from a breakdown is used to change the relative weight of the 633
points of the estimator to the left of it. Therefore, based on the observations obtained 634
from empirical data, Dirichlet distribution is chosen as a reasonable prior for our model 635
that estimates the capacity distribution. 636
The effects of the distribution parameters on the shape of estimated capacity 637
curves have shown to have the utmost importance. For instance, parameter directly 638
depends on the number of breakdown and censored observations. As a rule of thumb, the 639
larger it is, a better data set (better representative information) we have for the given 640
section. The parameter for the Dirichlet prior, on the other hand, defines the prior 641
weight on the traffic volume data. When the sample size of the observations is n , the 642
prior mass on (0, ) becomes n
. The larger it is, the smoother the Bayesian estimator 643
becomes compared to the Kaplan-Meier estimator. 644
Finally, the capacity intervals for the breakdown probabilities of ramp sections 645
make it possible for a traffic engineer/planner to choose the proper pairs of the 646
breakdown probability and traffic capacity. Freeway control strategies such as ramp 647
metering that heavily depend on the accurate estimation of the section capacities can be 648
improved by using this methodology. The temporal changes in the freeway capacity can 649
also be studied regarding the weather conditions, seasonal demand, and daylight/darkness 650
conditions. 651
652
ACKNOWLEDGMENTS 653
654 We would like to acknowledge California DOT for providing us with the PEMS 655
data. 656
657
REFERENCES 658
1. Geistefeldt, J., and Brilon, W., “A Comparative Assessment of Stochastic Capacity 659
Estimation Methods”, Transportation and Traffic Theory 2009, Edited by W.H.K. 660
Lam, S.C. Wong and H. K. Lo, pp. 583-602, Springer, 2009. 661
2. Lorenz, M. and Elefteriadou, L., “A Probabilistic Approach to Defining Freeway 662
Capacity and Breakdown”, Proceedings of the 4th International Symposium on 663
Highway Capacity, pp. 84-95, TRB-Circular-E-C018, Transportation Research 664
Board, Washington D. C, 2000. 665
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
26
3. Lorenz, M., and Elefteriadou, L., “Defining Highway Capacity as a Function of the 666
Breakdown Probability”, Transportation Research Record, Volume 1776, pp. 43-51, 667
2001. 668
4. Kim, J., Elefteriadou, L., “Estimation of Capacity of Two-Lane Two-Way Highways 669
Using Simulation Model”, ASCE Journal of Transportation Engineering, Volume 670
136-1, 2010. 671
5. Persaud, B., Yagar, S., and Brownlee, R., “Exploration of the Breakdown 672
Phenomenon in Freeway Traffic”, Transportation Research Record, Volume 1634, 673
pp. 64-69, 1998. 674
6. Chung, K., Rudjanakanoknad, J., and Cassidy, M.J., “Relation between Traffic 675
Density and Capacity Drop at Three Freeway Bottlenecks”, Transportation Research 676
Part B, Volume 41-1, p. 82-95, 2007. 677
7. Hall, F.L. and Agyemang-Duah, K., “Freeway Capacity Drop and the Definition of 678
Capacity”, Transportation Research Record, Volume 1320, pp. 91-98, 1991. 679
8. Cassidy, M.J., and Bertini, R.L., “Some Traffic Features at Freeway Bottlenecks”, 680
Transportation Research Part B, Volume 33, pp. 25-42, 1999. 681
9. Smilowitz, K., Daganzo, C.F., Cassidy, M.J. and Bertini, R.L., “Some Observations 682
of Highway Traffic in Long Queues”, Transportation Research Record, Volume 683
1678, pp. 225-233, 1999. 684
10. Cassidy, M.J, and May, A.D., “Proposed Analytical Technique for Estimating 685
Capacity and Level of Service of Major Freeway Weaving Sections”, Transportation 686
Research Record, Volume 1320, pp. 99-109, 1991. 687
11. Wang, M., Cassidy, M.J., Chang, P. and May, A.D., “Evaluating the Capacity of 688
Freeway Weaving Sections”, ASCE Journal of Transportation Engineering, Volume 689
119-3, pp. 360-384, 1993. 690
12. Lertworawnich, P. and Elefteriadou, L., “Capacity Estimation for Type B Weaving 691
Areas Based on Gap Acceptance”, Transportation Research Record, Volume 1776, 692
pp. 24-34, 2001. 693
13. Awad, W.H., “Neural Networks Model to Estimate Traffic Capacity for Weaving 694
Segments”, Proceedings of the 4th
International Symposium on Uncertainty 695
Modelling and Analysis, 2003. 696
14. Yeon J., Hernandez S., and Elefteriadou, L., “Differences in Freeway Capacity by 697
Day of the Week, Time of Day, and Segment Type”, Presented at the 86th
Meeting of 698
the Transportation Research Board, Washington D.C., 2007. 699
15. Kuhne, R., Mahnke, R., Lubashevsky, I., and Kaupuzs, J., “Probabilistic Description 700
of Traffic Breakdowns”, The American Physical Society, Physical Review E, Volume 701
65, 2002. 702
16. Kuhne, R., and Mahnke, R., “Controlling Traffic Breakdowns”, Proceedings of the 703
16th International Symposium on Transportation and Traffic Theory, pp. 229-244, 704
2005. 705
17. Kuhne, R., “Application of Probabilistic Traffic Pattern Analysis to Capacity”, Siam 706
Conference, 2001. 707
18. Uchida, K., and Munehiro K., “Impact of Stochastic Traffic Capacity on Travel 708
Time in Road Network”, Presented at the 89th Meeting of the Transportation 709
Research Board, Washington D.C., 2010. 710
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
27
19. Jun, Y., Hualiang, T., Heng, W., and Siji, H., “Estimating Roadway Capacity Using 711
the Simultaneous Spline Regression Model”, Journal of Transportation Systems 712
Engineering and Information Technology, Volume 9-1, pp. 87-98. 713
20. Yao, J., Rakha, H., Hualiang, T., Kwigizile, V., and Kaseko, M., “Estimating 714
Highway Capacity Considering Two-Regime Models”, ASCE Journal of 715
Transportation Engineering, Volume 135-9, 2009. 716
21. Banks, J.H., “New Approach to Bottleneck Capacity Analysis: Final Report, 717
California PATH Research Report, 1055-1425, UCB-ITS-PRR-2006-13, 2006. 718
22. Banks, J.H., “Effect of Time Gaps and Lane Flow Distributions on Freeway 719
Bottleneck Capacity”, Presented at the 85th
Annual Meeting of the Transportation 720
Research Board, 2006. 721
23. Banks, J.H., “Effect of Site and Population Characteristics on Freeway Bottleneck 722
Capacity”, Presented at the 86th
Annual Meeting of Transportation Research Board, 723
2007. 724
24. Banks, J.H., “Review of Empirical Research on Congested Freeway Flow”, 725
Transportation Research Record, Volume 1802, pp. 225-232, 2002. 726
25. Polus, A., and Pollatschek, M.A., “Stochastic Nature of Freeway Capacity and Its 727
Estimation”, Canadian Journal of Civil Engineering, Volume 29, pp. 842-852, 2002. 728
26. Laval, J.A., “Stochastic Processes of Moving Bottlenecks: Approximate Formulas for 729
Highway Capacity”, Transportation Research Record, Volume 1988, pp. 86-91, 2006. 730
27. Brilon, W., Geistefeldt, J. and Regler, M., “Reliability of Freeway Traffic Flow: A 731
Stochastic Concept of Capacity”, Proceedings of the 16th
International Symposium on 732
Transportation and Traffic Theory, pp. 125-144, 2005. 733
28. Brilon, W., Geistefeldt, J., and Zurlinden, H., “Implementing the Concept of 734
Reliability for Highway Capacity Analysis”, Transportation Research Record, 735
Volume 2027, pp. 1-8, 2008. 736
29. van Toorenburg, J., “Praktijwaarden voor de capaciteit”, Rijkswaterstaatdienst 737
Verkeerskunde, Rotterdam, 1986. 738
30. Minderhoud, M.M., Botma, H., and Bovy, P.H.L., “Assessment of Roadway Capacity 739
Estimation Methods”, Transportation Research Record, Volume 1572, pp. 59-67, 740
1997. 741
31. Regler, M., “Freeway Traffic Flow and Capacity”, Master Thesis, Institute for 742
Transportation and Traffic Engineering, Bochum University, 2004. 743
32. Geistefeldt, J., “Consistency of Stochastic Capacity Estimations”, Presented at the 744
89th
Meeting of the Transportation Research Board, Washington D.C., 2010. 745
33. Ozbay K, and Ozguven E.E., “A Comparative Methodology for Estimating the 746
Capacity of a Freeway Section”, Proceedings of the 10th International IEEE 747
Conference on Intelligent Transportation Systems, Seattle, WA, USA, Sept. 30 - Oct. 748
3, 2007. 749
34. Ozguven, E.E., and Ozbay, K., “A Nonparametric Bayesian Estimation of Freeway 750
Capacity Distribution from Censored Observations”, Transportation Research 751
Record, Volume 2061, pp. 20-29, 2008. 752
35. Susarla, V., and Ryzin, J.V., “Nonparametric Bayesian Estimation of Survival Curves 753
from Incomplete Observations”, Journal of the American Statistical Association, 754
Volume 71-356, pp. 897-902, 1976. 755
TRB 2011 Annual Meeting Paper revised from original submittal.
Ozguven E. E., Ozbay K.
28
36. Ferguson, T.S., and Phadia, E.G., “Bayesian Nonparametric Estimation Based on 756
Censored Data”, The Annals of Statistics, Volume 7, pp. 163-186, 1979. 757
37. Kaplan, E.L., and Meier, P, “Nonparametric Estimation from Incomplete 758
Observations”, Journal of the American Statistical Association, Volume 53, pp. 457-759
481, 1958. 760
38. Teh, Y.W., Jordan, M.I., Beal, M.J., and Blei, D.M., “Hierarchical Dirichlet 761
Processes”, Journal of the American Statistical Association, Theory and Methods, 762
Volume 101-476, 2006. 763
39. Susarla, V., and Ryzin, J.V., “Large Sample Theory for a Bayesian Nonparametric 764
Survival Curve Estimator Based on Censored Samples”, The Annals of Statistics, 765
Volume 6-4, pp. 755-768, 1978. 766
40. PEMS Database, http://pems.dot.ca.gov. 767
41. Kutz, M., Handbook of Transportation Engineering, McGraw-Hill, 2004. 768
42. Efron, B., “Censored Data and the Bootstrap”, Journal of the American Statistical 769
Association, Volume 76, Issue 374, 1981. 770
TRB 2011 Annual Meeting Paper revised from original submittal.