Available online at www.sciencedirect.

com

Procedia Social and Behavioral Sciences 17 (2011) 12–26

19th International Symposium on Transportation and Traffic Theory

Capacity Drops at Merges: an endogenous model

Ludovic Leclercqa,*, Jorge A. Lavalb, Nicolas Chiabauta
a
LICIT, IFSTTAR / ENTPE, Université de Lyon, Rue Maurice Audin, 69518 Vaulx-en-Velin Cedex, France
b
School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, GA, USA

Abstract

The Newell-Daganzo merge model is not only very simple but also accurately reproduces experimental findings. However, the
capacity downstream of the merge is an exogenous variable in the model. This is a serious limitation for merges that behave as
active bottlenecks because their downstream capacity is a direct consequence of the merging behavior. This paper proposes an
analytical model that extends the Newell-Daganzo model by incorporating, endogenously, the capacity drop related to the
merging process. Two cases are investigated depending on the traffic states on the on-ramp. The model properties are analyzed
and a sensitivity analysis is performed to quantify the relative contribution of the each parameter in the capacity drop. Finally, the
extended Newell-Daganzo model is validated with experimental data coming from an active merge bottleneck on the M6 freeway
in UK.
© 2011 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.

Keywords: Capacity drop, merge, merge ratio, kinematic wave model, active bottleneck

1. Introduction

Traffic flow behavior at freeway merges has been extensively investigated in the literature. One of the simplest
models for merges was originally proposed in (Newell, 1982) and later formalized in (Daganzo, 1995). This model,
referred to as ND model in this paper, postulates that flows on both incoming roads share the available downstream
supply following a specific allocation scheme, see Fig. 1. Notably, when the main road is congested (high demand),
the on-ramp may be either (i) in free-flow (low demand) or (ii) congested (high demand). In case (i), vehicles on the
on-ramp force their way and all the demand succeeds in merging. In case (ii) the merging behavior is mostly
deterministic and dictated by a fixed merge ratio. This simple model has been verified repeatedly to accord
surprisingly well with experimental findings (Troutbeck, 2002; Cassidy and Ahn, 2005; Bar-Gera and Ahn, 2010).
The main limitation of the ND model is that the downstream capacity is exogenous. This prevents any
meaningful modeling of the capacity drop phenomenon or the effects of the merging section length. This is
particularly troublesome when the merge behaves as an active bottleneck. Indeed, capacity drops between 10 to 30
% are commonly observed in that case (Elefteriadou et al., 1995; Persaud et al., 1998; Kerner, 2002; Cassidy and

* Corresponding author. Tel.: +33-472047716; fax: +33-472047712.

E-mail address: [email protected].

1877–0428 © 2011 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
doi:10.1016/j.sbspro.2011.04.505
 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26 13

Bertini, 1999; Sarvi et al., 2007; Chung et al., 2007; Yi and Mulinazzi, 2007). Lane-changing, low-speed inserting
vehicles and heterogeneous lane behavior have been suspected to induce the capacity drop (Cassidy and
Rudjanakanoknad, 2005; Cassidy and Ahn, 2005; Laval et al., 2005, Treiber et al., 2006, Laval and Daganzo, 2006).
Indeed, these phenomena generate variations between over and mid-saturated traffic states at merges (Mauch and
Cassidy, 2002; Ahn and Cassidy, 2007; Laval et al., 2009; Li et al., 2010) that prevent to reach the full freeway
capacity.
Several models have been proposed to account for capacity drops, but with the exception of (Laval and Daganzo,
2006), it is treated exogenously. For example, (Koshi et al. 1983; Hall and Hall, 1990) propose models that are
mainly descriptive and based on fundamental diagrams with reverse lambda shape; (Evans et al., 2001; Kerner, 2000
; 2004) postulate stochastic approaches; and (Siegel et al., 2009 ; 2009b) assume second order models derived from
(Aw and Rascle, 2000) and (Greenberg, 2001) works. However, none of these models propose an explicit
formulation of the relationship between local traffic interactions related to lane-changing and their global impact on
the capacity. Furthermore, most of these models are hard to implement and to calibrate in practice.
The aim of this paper is to overcome these issues by introducing, in the ND model, an endogenous analytical
description of the merge bottleneck activation and the ensuing capacity drop. Basic idea is to incorporate in this
model the physical mechanism unveiled in (Laval and Daganzo, 2006) pertaining to the effects of the bounded
acceleration of merging vehicles. Toward this end, section 2 of the paper presents the analytical model that estimates
the capacity drop with respect to the demand on the on-ramp and the different model parameters. For mathematical
tractability, this model is not fully realistic because both incoming roads are assumed single-lane. However, it will
significantly improve the global understanding of how the merging mechanism contributes to the capacity drop. A
sensitivity analysis is performed in section 3 to precisely quantify the contribution of the different parameters. This
will provide some interesting insights on how road design can improve merge capacity. Finally, the extended ND
model will be validated with experimental data coming from a merge on the M6 freeway in UK in section 4. To this
end, an estimation method will be presented to partly overcome the one-lane limitation of the model. Finally, section
5 presents a discussion.

Fig. 1: Merge diagram for the Newell-Daganzo (ND) model

2. An analytical model for capacity drop

In this section, a simple analytical model that accounts for the bounded acceleration of inserting maneuvers is
presented. Two cases will be studied separately depending on the state of the on-ramp, which can either be
congested (right top quadrant of ND model) or not (right bottom quadrant), see Fig. 1. The demand on main road is
always high enough to encounter congestion.
14 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26

2.1. Congested on-ramp

Consider a merge with two one-lane incoming roads as in Fig. 2a. The main road is labeled 1 and the on-ramp is
labeled 0. The length of the insertion section is L. Traffic on each road is described by the kinematic wave model
(Lighthill and Whitham 1955; Richards 1956) and a triangular fundamental diagram with free-flow speed u, wave
speed w, and jam density N. The capacity on one-lane is equal to Q=wuN/(w+u). The upstream demands are denoted
Oi (i=0,1) and the mean flows just before the beginning of the insertion section (x=0) are noted qi. The merge always
behaves as an active bottleneck, i.e. the downstream flow q0+q1 is never constrained by the downstream supply P.
Vehicles from the on-ramp are assumed to insert themselves in the main road with speed v0 and then accelerate at
a constant rate a until they reach the free-flow speed, see Fig. 2b. Both v0 and a are assumed identical for all
vehicles. This seems reasonable as the on-ramp is congested and thus the traffic states are quite homogeneous.
Vehicle i is inserting at time ti. The time headway between two consecutive insertions is denoted hi=ti+1-ti. Such
headways follow an unknown distribution H(h0,s) with mean h0 and standard deviation s. When s is equal to 0,
vehicles regularly insert every h0, as illustrated in Fig. 2b. After their insertion, vehicles from the on-ramp behave as
moving bottlenecks (Newell, 1998; Lebacque et al., 1998; Munoz and Daganzo, 2002; Leclercq et al., 2004). Thus,
these moving obstructions constrain the flow on the main road and create voids in front of them. It was shown in
(Laval and Daganzo, 2006) that these moving obstructions are responsible for the capacity drop. Two cases will be
distinguished in the present work: L=0 and L>0.

Fig. 2: (a) sketch of the merge (b) Inserting process when L=0

When L=0, all insertions take place at x=0, see Fig. 2a. Let Ni be the cumulative number of vehicles that have
crossed this location by time ti. At a large time scale, the mean total flow downstream of the merge, i.e. the effective
capacity C, is the same whatever the location is. Thus, C can be simply expressed at x=0 with respect to Ni:
n n
C = ¦ ( N i +1  N i ) / ¦ hi with n o +f (1)
i =1 i =1

Variational theory (Daganzo, 2005) provides a convenient way to calculate the increase in N between ti and ti+1,
i.e. along the horizontal path from A to B, see Fig. 2b. Indeed, this increase is the same as on the alternative path
AoCoB that first follows the moving bottleneck trajectory and then reaches B along a characteristic with slope w.
From A to C, the passing rate is equal to 0 because no vehicle can overpass the moving obstruction. From C to B,
the passing rate is equal to wN because the path slope is w. Thus, the increase in N between ti and ti+1 is equal to
wN(hi-W(hi)), where W(hi) represents the time duration between points A and C, see Fig. 2b and (Leclercq, 2005) for
more details. It can be shown that the effective capacity C can be obtained with:
§ n n
· w + v0 1
C = wN ¨ 1  ¦ W (hi ) / ¦ hi and W (hi ) =  + (w + v0 )2 + 2awhi (2)
© i =1 i =1 ¹̧ a a
 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26 15

The law of large numbers tells us that:

n n
(1 / n)¦ hi o E (hi ) and (1 / n)¦ (hi ) o E ( (hi )) (3)
i =1 i =1

where E(x) is the mathematical expectation of x. E(hi) is equal to h0. As the distribution H of hi is unknown, E(W(hi))
cannot be analytically derived. However, one can estimate its second-order approximation (Oehlert, 1992):

(4)

Thus, the effective capacity C is equal to:

(5)

This effective capacity is shared by the two incoming flows q0 and q1. As both upstream roads are congested, the
merge ratio D holds and q0=Dq1 and the effective capacity can also be expressed as:
C = q0 + q1 = (1 + 1 / ) q0 (6)

Note that q0 and v0 are related by the fundamental diagram, i.e. v0=wq0/(wN-q0), and that h0=1/q0. Thus, by
combining (5) and (6), we obtain an equation in q0 with respect to the following parameters: a, w, N, D and s:

§ w 3N 2 wN w 4N 2 as 2 w 3N · 1·
2aw §
wN + q0 ¨  + + ¸  ¨1 + q0 = 0 (7)
¨© a(wN  q0 ) a (wN  q0 )2 q0 2 w 4N 2 (wN  q0 )2 + 2 aw q0
( )
3/2
¸¹ © D ¹̧

Unfortunately it is impossible to derive the explicit formulation of q0 with respect to these parameters. However,
we can easily compute the q0 values for any given sets of parameters. We propose the indicator c to quantify the
relative capacity drop, i.e. the complement of the ratio between the effective capacity C and the capacity Q given by
the fundamental diagram:
C § 1 · q0
c = 1 = 1  ¨1 + (8)
Q © D ¹̧ Q
Note that capacity drop here cannot directly be compared with experimental values found in the literature.
Indeed, we have chosen a fixed reference, Q, to calculate the capacity drop. In reality, the capacity drop is often
defined in reference to the maximal flow observed just before the capacity drop, which is always lower than Q.
Further in the paper we will evaluate how a, w, N, D and s influence c by performing a sensitivity analysis.
Now we consider the case of a merge where the insertion section has a spatial extension (L>0). In this case,
vehicle i's insertion takes place at a random location xi [0,L]. The distribution X of xi is assumed to be uniform, as
suggested by the empirical evidence in (Daamen et al., 2010). Indeed, these authors have computed merge location
histograms at a merge in Netherlands. This distribution appears to be piecewise uniform in congestion, see figure 6
in that reference. We have performed extensive tests that demonstrate that a piecewise uniform distribution leads to
the same results in terms of capacity drop as a uniform distribution.
To simplify the exposition, vehicles are first supposed to enter the mainline every h0, i.e. s=0, see Fig. 3a. A
supplementary assumption is made here: we neglect the effect that the void just downstream of a moving bottleneck
can have on its own trajectory by making it react to downstream congestion waves carrying lower speeds. This
implies that a moving bottleneck disappears when it crosses such waves; see square dots in Fig. 3a. We have
confirmed that this assumption does not introduce significant error when L is not too high (roughly below 300 m),
and therefore appears to be a valid assumption in our context. Note that this moving bottlenecks behavior is close to
the relaxation process described in (Laval and Leclercq, 2008) for lane-changers.
16 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26

Fig. 3: (a) Inserting process when L>0 (b) Study of the standard deviation of H’

Let t’i be the time when the wave coming from (ti,xi) reaches x=0 and t’(i) the ordered series built from the
realizations of t’i, see Fig. 3a. The effective capacity C can still be calculated at x=0 with (1) by focusing on the
increase in N between t’(i) and t’(i+1), i.e. between points A and B in Fig. 3a. This increase can be calculated along
the alternative path AoCoDoB in figure. Note here a crucial property: if the moving boundary condition from C
to D is translated by the vector CA, we can see another alternative path AoD’oB where the increase in N is
equivalent, see Fig. 3a. Thus, the problem appears to be equivalent to the previous one with L=0 by substituting H
by H’, the latter being defined as the distribution of h’i= t’(i+1)-t’(i). The distributions H and H’ have the same mean
h0. We only need to determine s’ the standard deviation of H’ and the effective capacity C will be obtained by
substituting s by s’ in (5).
The computation of s’ is a quite challenging task because there is no obvious relation in most cases between ti
and t’(i) due to the ordering process. A particular case can first be studied. When L<wh0, the order of t’i is not
modified as every wave emitted at time ti reaches x=0 before ti+1. Thus, h’i= t’(i+1)-t’(i)= t’i+1-t’i=h0+(xi+1-xi)/w in this
case. Inserting positions of two consecutive vehicles may reasonably be considered as independent. The standard
deviation s’ of H’ is then given by:
s '2 = V (h 'i ) = (1 / w2 )V ( xi +1  xi ) = L2 6w2 (9)

where V(x) is the variance of x. When L is very large (formally tends to infinity), it can be demonstrated that H’
follows an exponential distribution. Thus, its standard deviation becomes equal to its mean h0. When L is higher
than wh0 but not too large. We have performed extensive numerical analysis of s’ with respect to L for different
values of h0 and w. It turns out that in all cases the best fit is obtained with the following formula:
s '( L ) = ( bL + c ) ( L + d ) (10)

Fig. 3b presents the fitting result for h0=5 s and w=5.38 m/s. The r-square is higher than 0.99. Similar fits are
observed for a wide range of values for h0 and w. Interestingly, it is possible to determine the analytical formulations
of b, c and d. First, as s’ tends to h0 when L tends to infinity, b should be equal to h0. Second, we can reasonably
impose that s’ and its first derivative be continuous at L=wh0, see Fig. 3b. Thus, the final expression of s’ is:
­ L 6w if L < wh0
°
s '(L) = ® (11)
(
°̄h0 L  wh0 6 ) (L + ( ) )
6  2 wh0 if L t wh0

We can now relax the assumption s=0 when L>0. Let H’’ be the distribution of the time gaps h’’i between the
start of two consecutive moving obstructions at x=0. H’’ has the same mean h0 as H. Its standard deviation s” is
driven by two processes, one related to the distribution of the inserting positions and the other related to the
distribution of the inserting times. If we suppose that these two processes are independent, s” can be estimated by:
 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26 17

s '' = s 2 + s '2 (12)

To summarize, it appears that when both incoming roads are congested, (7) and (8) can be applied to estimate the
capacity drop in the below mentioned four cases. One only has to substitute s in (7) by the right expressions:
1. sm0, point merge with constant time gaps between two successive insertions (L=0, s=0);
2. sms, point merge with distributed time gaps between two successive insertions (L=0, s>0);
3. sms’ given by (11), extended merge with constant time gaps between two successive insertions (L>0,s=0);
4. sms” given by (12), extended merge with distributed time gaps between two successive insertions ((L>0,s>0).

2.2. Uncongested on-ramp

In this section, we focus on the right bottom part of the ND merge diagram in Fig 1. Here, O0 is low enough to
prevent congestion on the on-ramp and therefore q0=O0. Thus, speeds on the on-ramp are no longer constrained. We
assume then that inserting vehicles will adopt the mean main road speed v1 before accelerating. The inserting
process remains the same as in the previous section. The effective capacity C is given by substituting v0 by v1 in (5).
Note that h0 is now equal to 1/O0 and becomes an additional parameter. Note too that v1 and q1 are related by the
fundamental diagram, i.e. v1=wq1/(wN-q1). Since C is also equal to q1+O0, we obtain the following equation in q1:

1 § q1 · 1 w 2N 1 w 4N 2 2aw aw 2 s 2
wN ¨© 1 + O ¹̧ = O + a ( wN  q )  a +
O0
+ (13)
0 0 1 ( wN  q1 ) 2
(w N
4 2
( wN  q1 )2 + 2aw O0 )
3/2

Interestingly, the above expression can be greatly simplified when s=0 by defining q’1=wN-q1 and E=2w3N2/a:

0 = q '1( 3/2 ) ( q '1 )( q '1 ) (14)

This expression shows that when E increases, q’1 increases and q1 decreases for a given value of O0. We fail to
determine a convincing physical interpretation of E but interestingly, E increases when a decreases. This
corroborates an intuitive result: the capacity drop is higher when the acceleration rate is lower. To complete this
study, we still have to derive the expression of s’ when L>0 because (11) is no longer applicable. Indeed, figure 6 in
(Daamen et al., 2010) shows that the merge location histogram in free-flow fits a normal distribution rather than a
uniform one. Furthermore, this figure suggests that the mean inserting position Lm is close to L/4. Thus, we assume
that X follows a normal distribution with mean Lm and standard deviation Lm/2.57. The standard deviation has been
adjusted so that 99.5% of the insertions take place between 0 and L. The expression of s’ can then be obtained using
a similar method as in previous sections:
­ L 2.57 8 w if L < 2.7 wh0
°
s '(L) = ® (15)
(
°̄h0 L  2.7 wh0 2.57 8
2
) ( L + ( 2.57 ) )
8  2.7 wh0 if L t 2.7 wh0

Note that the threshold from when the ordering process of t’i influences the standard deviation of h’i is now
higher than wh0. This is because most insertions are generated at the beginning of the insertion section, i.e. roughly
between 0 and L/2. By combining (15)-(13) or (12)-(15)-(13) it is possible to determine the value of q1 with respect
to O0 in all four case studies. The relative capacity drop c is then given by:
c =1 ( q1 + 0 ) Q (16)
18 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26

3. Analysis of the model properties

3.1. Results for a reference scenario

Both analytical models introduced previously are applied for the following reference scenario: w=19.4 km/h,
u=115 km/h, N=145 veh/km, a=2 m/s2, D=0.76. The theoretical capacity Q given by the fundamental diagram is then
equal to 2400 veh/h. The length L is set to 0 m (point merge) or to 160 m (extended merge). The standard deviation
s is either equal to 0 s (constant time gaps) or 2 s (distributed time gaps). Fig. 4a presents the evolution of c with
respect to O0. For O0 values lower than 100 veh/h, no capacity drop occurs. The relative capacity drop then gradually
increases until it reaches its maximum value when the on-ramp is fully congested. The threshold value for O0 is
about 600 veh/h. When O0 exceeds this threshold, c remains constant. Note that the maximal value for c is very high
(up to 46%). This is because the relative capacity drop is defined with respect to the theoretical reachable capacity Q
and because there is only one lane downstream.
For lower values of O0, neither the spatial extension of the merge nor the standard deviation s have a noticeable
influence on the capacity drop. When O0 increases, the four curves begin to diverge. At the end when the on-ramp is
fully congested, it appears that the relative capacity drop is reduced for 46% to 40% when L changes from 0 to
160 m and s=0. This corresponds to an increase of the effective capacity from 1310 to 1450 veh/h. Thus, the spatial
extension of the merge significantly improves the insertion process.
Fig. 4b presents the same results but plotted in the merge diagram. Note that when the on-ramp is uncongested,
the inserting flow q0 is always equal to O0. The diagonal line in this figure corresponds to the theoretical capacity Q
derived from the fundamental diagram. It helps to better appreciate the absolute value of the capacity drop with
respect to the flow on the on-ramp.
(a) q [veh/h] (b)
c L=0,s=0 0
0.5
L=0,s>0 1000 ted L=0,s=0
L>0,s=0 ges
n L=0,s>0
0.4 L>0,s>0 co
s L>0,s=0
800 pi
m L>0,s>0
ra
0.3 on
600
th
eo
0.2 rit
on ramp is congested 400 i ca
l
ca
pa
cit
0.1 200 yQ

0
[veh/h] q1 [veh/h]
0 0
0 200 400 600 800 1000 550 1000 1500 2000 2400

Fig. 4: (a) Relative capacity drop with respect to on-ramp demand (b) Traffic states in ND framework

3.2. Sensitivity analysis

The sensitivity of the relative capacity drop c to the different parameters a, w, N, D, L and s can be evaluated with
(7) and (13). We first focus on the acceleration rate a for the reference scenario when L>0 and s=0. Fig. 5a presents
the evolution of c with respect to O0 for increasing values of a from 0.5 to 3 m/s2. This parameter appears to have a
high influence. For example, when the on-ramp is congested, c is reduced from 0.49 to 0.34 when the on-ramp is
congested and a varies from 1 to 3 m/s2. This corresponds to an increase of the effective capacity from 1220 to 1580
veh/h (+23%). When the on-ramp is uncongested and O0=300 veh/h, c decreases from 0.32 to 0.16 for the same
variation in a. The effective capacity thus increases from 1630 to 2020 veh/h (+22%). It is clear that merges should
preferentially be implemented where vehicles could optimally accelerate, i.e. no uphills, adequate sight distance,…
Fig. 5a also highlights that both capacity drop beginnings and full congestion appearance on the on-ramp are
 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26 19

observed for higher O0 values when a increases. Fig. 5b shows the evolution of the effective capacity C in the merge
diagram for a between 0.5 and 3 m/s2. This confirms that the capacity drop is highly influenced by the acceleration.
Indeed, the lower the acceleration is the farer the effective capacity is from the theoretical capacity line.
(a) q0 [veh/h] (b)
c a=0.5
1000
0.5 a=1.0

the
a=1.5

or
a = 2. .0
800

iti
a=2.0

a= =2 1.5
3. 5
a = .0

ca
0.4

0
a 1

lc
a=2.5

a=

ap
a=3.0 600

acit
0.3

a=

y
0.

Q
5
400
0.2
on ramp is congested

0.1 200

0
[veh/h] q1 [veh/h]
0 0
0 200 400 600 800 1000 200 500 1000 1500 2000 2400
Fig. 5: Influence of the acceleration on the capacity drop (a) merge diagram representation (b) relative capacity drop with respect to O0

The sensitivity analysis is now extended to all parameters when the on-ramp is congested. Free-flow situations
are disregarded because they mostly lead to similar conclusions. Each parameter is first separately tested for a
typical range of values, see Fig. 6. The four identified cases for the insertion process are considered and the
reference value for each parameter is pointed out by a dotted line in each subplot. Fig. 6a first reinforces that the
acceleration rate is the most influent parameter. The wave speed and the jam density have also a significant impact,
see Fig. 6b-c. However, note that a wide range of values have been tested for these parameters. In reality, w and N
are known to vary little for a given freeway. Furthermore, these parameters also modify the capacity Q of the main
road. Thus, when w increases (respectively N), Q also increases and c then increases even if the effective capacity C
is constant.
(a) (b) (c)
c c

0.5 0.5 0.5

0.4 0.4 0.4

0.3 0.3 0.3

L=0,s=0
0.2 L=0,s>0 0.2 0.2
L>0,s=0
0.1 0.1 0.1
L>0,s>0
a [m/s2] w [km/h] [veh/km]
0 0 0
0.5 1 1.5 2 2.5 3 15 20 25 120 140 160 180 200
(d) (e) (f)
c c c

0.5 0.5 0.5

0.4 0.4 0.4

0.3 0.3 0.3

0.2 0.2 0.2

L>0,s=0 L=0,s>0
0.1 0.1 0.1
L>0,s>0 L>0,s>0
L [m] s [s]
0 0 0
0.5 1 1.5 0 50 100 150 200 1 2 3 4 5 6
Fig. 6: Sensitivity analysis when on-ramp is congested; influence of (a) the acceleration (b) the wave speed (c) the jam density (d) the merge
priority ratio (e) the length of the insertion section (f) the standard deviation of H.
20 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26

The merge priority ratio seems to have almost no influence on capacity drop, see Fig. 6d. This is quite surprising
because when D increases the inserting flow q0 increases and so do the number of moving bottlenecks. This would
have to result in a drop of the effective capacity. However, the analytical model sheds some light on the global
mechanism. Indeed, increasing q0 increases the number of inserting vehicle but also their speed v0 because traffic
conditions improve on the on-ramp. Both phenomena compensate each other and the capacity drop is almost steady
regardless of the merge ratio. Thus, all traffic states fall on the same capacity line in the merge diagram.
Fig. 6e focuses on the influence of the length of the insertion section. The figure shows that increasing L reduces
the relative capacity drop and then improves the effective capacity. It is noticeable that most of the benefit is
obtained before L reaches 100 m. If this result could be generalized with more than one-lane on the main road, it
would be an interesting insight for road design. Subplots a to d also highlight the benefits of an extended merge by
comparing plots when L=0 and L=160 m (L>0). Note in Fig. 6e that when s=0 and L tends to 0, the relative capacity
drop converges to the value obtained when L=0 (gray dot). Thus, the transition between L=0 and L>0 is smooth.
Fig. 6f shows the influence of the standard deviation of H when L=0 and L>0. It appears that increasing s reduces
c. Indeed, randomness in inserting times for on-ramp vehicles sometimes generates larger time headways where the
flow on main road is able to reach higher values. Thus, the effective capacity increases. This also explains why
increasing L reduces the capacity drop. Indeed, we have previously demonstrated that a random distribution of
inserting positions is equivalent to introducing randomness in H.
The last part of the sensitivity analysis consists in studying the joint influence of the different parameters when
the on-ramp is congested. We disregard D because it has no influence on the capacity drop and fixe s to 0. Fig. 7a
shows the joint influence of a and L on the relative capacity drop c. Dots corresponds to the values provided by the
analytical model when L>0 and s=0. It appears that a polynomial regression with degrees not higher than 3 perfectly
fits dots (r-square equal to 0.9994). Furthermore, the joint terms of the regression are negligible because r is not
significantly reduced when they are omitted (r turns to 0.9993). We finally obtain the following expression:
c = 0.774  0.331a + 0.122a2  1.85.102 a3  6.76.104 L + 1.73.106 L2 (17)

(a) (b)

c c

0.6 0.6
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
0.5 0.5
1 15 1
50 17
100 1.5 19 1.5
150 2 a [m/s2] 21 2 a [m/s2]
L [m] w [km/h] 23
2.5 2.5
Fig. 7: Multiple regression (a) c with respect to a and L (b) c with respect to a and w.

Thus, we can conclude that a and L have independent contributions to the relative capacity drop. Similar studies
are performed for a and w, see Fig. 7b, and for a and N. Polynomial regressions with no joint terms provide accurate
fits with r equal respectively to 0.9993 and 0.9994. Thus, all parameters look independent. It is then possible to
combine all the obtained regressions into one:
c = 0.402 0.332a + 0.122a2 1.85.10 2 a3 6.76.10 4 L + 1.73.10 6 L2 + 6.84.10 2 w 3.12.10 3 w2 + 0.724 (18)
Where a is expressed in m/s2, L in m, w in m/s and N in veh/m. This global regression is surprisingly accurate: when
(18) is tested against model dots for the three planes (a,L), (a,w) and (a,N), the associated RMSE are respectively
3.4.10-3, 4.8.10-3 and 3.2.10-3. Note that (18) predicts the upper limit for the capacity drop because s=0, and provides
a ready to use extended ND model with an endogenous calculation of the capacity drop. Indeed, when the on-ramp
is congested, one only has to compare the downstream supply with the effective capacity derived from (18) to
 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26 21

determine the effective supply and then apply the merge ratio. Unfortunately, this formula is only valid for one-lane
on main road. We will see in the next section how we can overcome this limitation.

4. Model validation

4.1. Experimental data

Fig. 8: (a) sketch of the experimental site (b) speed evolution at detector M7088 (c) flow evolution at detectors M7092 and M7088 (d) flows by
lane at detector M7088.

To perform the model validation, we focus on a merge located on a southbound three-lane segment of the M6
highway near Manchester, UK. A sketch of the site is depicted in Fig. 8a. An aerial photograph can be consulted on
Google Earth® by looking for the following geographical coordinates: 53°25’14.85”N, 2°34’42.18”O. Regularly
staggered loop detectors can be found on this highway section. These detectors provide average flow, speed and
occupancy rate per lane every minute. Detector M7072 is located 1600 m downstream of the selected merge and
never encounters congestion. This confirms that the selected merge is really an active bottleneck. In this study, we
mainly focus on data from detectors M7092 and M7088 that are respectively located 250 m upstream and just
downstream of the merge, see Fig. 8a. Another detector M7092M is also useful because a drop in speed here
indicates that the on-ramp is fully congested. Unfortunately, this loop is far upstream of the beginning of the
merging section. Thus, it is not possible to deduce the inserting flow from these observations in a synchronized
manner with observations at M7088 and M7092. Inserting flows are then estimated by the difference between
22 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26

observations at M7088 and M7092 with a one-minute lag. This lag roughly represents the time that a vehicle needs
to travel from M7092 to M7088 in congestion.
All data from May 2006 is available. However, congestion is not observed every day nearby the merge. We have
then selected the morning peak hours for four days, i.e May, 2, 3, 9 and 16. Those days suffer the highest levels of
congestion. Fig. 8b-d present the observations for May, 3 rd between 6:00 and 10:00. Fig. 8b shows the evolution of
the speed at M7088, where one can identify the time when the freeway is congested, i.e between 6:45 and 9:30.
Combining this with information from M7092M, it is possible to determine when the on-ramp is in free-flow (light
congestion on the freeway) or in congestion (heavy congestion on the freeway). In that latter case, Fig. 8c clearly
shows stop-and-go waves that appear nearby the merge and that form oscillations.
When the ramp is uncongested, we select time periods when the flows observed at M7092 and M7088 are nearly
stable during at least 3 min, e.g. point C in Fig. 8c. When the ramp is congested, we focus on situations where flows
on each lane at M7088 are nearly equal (homogeneous lane distribution), e.g. points A 1, A2, B1 and B2 in Fig. 8d.
Such a choice will be justified in the next section. At the end of this data filtering process, we obtain couples of
insertions versus upstream flows that can be plotted in the merge diagram.

4.2. Comparison with the analytical model predictions

Fig. 9 presents the results of the experimental observations (void circles) for the four selected days. Note that the
other marks and the meaning of the grey shaded areas will be defined later in this section. Note also that the
fundamental diagram has already been calibrated on this site during congestion (Chiabaut et al, 2008): w=19.4 km/h
and N=145/km/lane. The free-flow speed u can easily be estimated as 115 km/h by looking at Fig. 8b. This is very
close to the speed limitation on UK motorways. The maximum capacity per lane given by the fundamental diagram
is thus Q=2400 veh/h.
q [veh/h] May 2, 2005 q [veh/h] May 3, 2005
0 0
Experimental observations
2000 Mean observation (Group A) 2000
Model (congested on ramp) A
Model (uncongested on ramp) =0.32 A
1500 1500 =0.3
C
1000 1000
B
B
C
500 500

0 q1 [veh/h] 0 q1 [veh/h]
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000
q0 [veh/h] May 5, 2005 q0 [veh/h] May 16, 2005

=0.39
2000 2000 =0.37
A A
1500 1500
C
C
B
1000 B 1000

500 500

0 q [veh/h] 0 q [veh/h]
0 1000 2000 3000 4000 5000 6000 1 0 1000 2000 3000 4000 5000 6000 1

Fig. 9: Comparison between experimental data and model estimations

 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26 23

Fig. 9 demonstrates one more time the relevance of the ND model. Indeed, when the on-ramp is congested, the
merging behavior respects a fixed priority ratio; see groups A and B on Fig. 9. When the on-ramp is uncongested,
traffic states fall into the right bottom quadrant of ND model; see group C. Notice here that a capacity drop is
observed in both cases as predicted by the analytical model. Note also that when the on-ramp is congested, traffic
states can be split in two groups. Group B corresponds to heavy congested situations correlated with appearances of
stop-and-go waves. Group A corresponds to regular congested situations.
To compare experimental data with the analytical model predictions, the one-lane limitation should be overcome.
To this end, we assume that the analytical model properly deals with the merging behavior between the on-ramp and
the shoulder lane. The challenge is to account for the lane flow distribution and discretionary lane-changing. We
first focus on situations when the on-ramp is congested. The merge ratio D represents the ratio between the inserting
flow q0 and the total upstream flow q1. The local merge ratio D’ refers to the ratio between q0 and the upstream flow
on the shoulder lane q1,1. The sensitivity analysis of the analytical model brings to light that D’ does not influence
the relative capacity drop c. Thus, even if D’ is unknown and potentially variable in time, it is possible to estimate
the downstream capacity on the shoulder lane C=(1-c)Q. With a=2 m/s2, s=2 s and L=160 m, the analytical model
leads to c=0.314 and C=1647 veh/h. When the lane flow distribution is homogeneous downstream of the merge, it is
easy to determine the total downstream capacity C’=3C. This explains why we have only selected, during
congestion on the on-ramp, experimental data with homogeneous lane flow distribution at M7080. These
observations can adequately be compared with the analytical model.
The model estimation is represented by a solid square in the four subplots of Fig. 9. Note that q0 and q1 are
derived from C’ using the observed merge ratio D. Fig. 9 highlights that the analytical result fit the average observed
traffic states of group A, i.e the relative error is always below 7%. This average corresponds to the solid black dots.
We can then conclude that the analytical model accurately estimates the regular congested situations for a congested
on-ramp when downstream lane distribution is homogeneous. However, Fig. 9 also shows that the analytical result
is far from group B observations. This is explained because the proposed model does not take into account stop-and-
go waves.
Let us now focus on situations when the on-ramp is uncongested. It is difficult to find downstream homogeneous
lane distributions in that latter case because the global freeway is not heavily congested. Most of the time, flows are
highly different on each lane, see Fig. 8d. The median lane can even be in free-flow while the shoulder lane is
congested. To compare experimental data with analytical results, we first determine the evolution of the capacity C
of the shoulder lane with respect to the on-ramp demand O0 using (13). Note that the analytical model in free-flow is
only valid when D’<1 because it implicitly assumes that the inserting flow is lower than the main road upstream
flow. This is most always the case when the main road has only one lane but this is not necessarily the case here.
Indeed, when the congestion is not severe on the freeways, a great proportion of vehicles on the shoulder lane may
switch to the center and the median lane. This increases the inserting flow. This limitation can be easily overcome
by considering that C(O0) has a lower-bound equal to the value calculated when the on-ramp is congested.
To derive the total capacity C’(O0) of the freeway from C(O0) we consider the two following borderline cases: (a)
all downstream lanes experience the same capacity restriction (high level of congestion) and (b) only the shoulder
lane experiences a capacity drop while center and median lanes are in free-flow (low level of congestion). The total
capacity is then either equal to (a) C’(O0)=3C(O0) or (b) C’(O0)=C(O0)+2Q. The region between these two curves
C’(O0) describes the capacity drop for all possible downstream lane flow distributions. This region is filled with light
gray in Fig. 9. It appears that almost all the experimental points from group C fall into the region predicted by the
analytical model. This model is therefore able to accurately define the range of possible observed capacity drop
when the on-ramp is in free-flow. Note that we have only selected stable traffic flow during at least 3 min in that
latter case because the analytical model only predicts mean capacity drop in time, see Fig. 8a.

5. Discussion

We have introduced an analytical model that predicts the capacity drop at an active merge bottleneck with respect
to the demand on the on-ramp. This model is quite simple while accounting for most of the key elements of the
merging mechanism. It predicts that the relative capacity drop c almost linearly increases with O0 when the on-ramp
is in free-flow, then reaches its maximal value and remains constant when the on-ramp is congested, see Fig. 4a and
Fig. 5a. This analytical model sheds some light on the relative influence of the different parameters involved in the
24 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26

merging process. The acceleration rate has the highest impact because low rate significantly reduces the available
capacity. This is not surprising but the insight here is that the capacity drop value can be directly derived from the
mean acceleration rate. The analytical model also predicts that the relative capacity drop is reduced when the length
of the insertion section increases. Note here that most of the reduction is observed before the length reaches 100 m.
Thus, longer insertion sections do not seem to significantly improve the capacity on freeways. This result has to be
confirmed experimentally but it is potentially valuable for road design. The last insight provided by the analytical
model is that the merge ratio does not influence the capacity drop when the main road has only one lane. This is not
obvious because higher merge ratios lead to higher inserting flows. But this increase is offset by a speed increase on
the on-ramp that reduces the perturbations caused by inserting vehicles. This property is convenient when estimating
capacity drops for freeways that have more than one-lane, see section 4. All the results synthesized here are certainly
useful to improve ramp-metering algorithms.
The first experimental results presented in this paper show that the analytical model is promising. The main
current limitation is that it needs to be completed with a lane flow distribution model that accounts for the mean
effect of discretionary lane changing when the main road has several lanes. However, it already provides an efficient
and elegant way to estimate the upper (when the freeway is uniformly congested) and lower (when congestion only
occurs on shoulder lane) bounds of the capacity drop. The authors currently investigate lane flow distribution at
merges. The second limitation is that the merge ratio has been fully considered as an exogenous parameter. This
hypothesis is reasonable as a first approximation as existing works show that this ratio mostly depends on the merge
design (Bar-Gera and Ahn, 2010). However, this should be more deeply investigated in the future because the
inserting process may also influence such a ratio when both upstream roads are congested. The last limitation of the
analytical model is that it does not reproduce stop-and-go waves when the on-ramp is congested but only the regular
capacity drop due the merging process. However, we found that oscillations do not appear due to the random
insertion process when vehicles are homogeneous and have the same acceleration rate. We can then conjecture that
drivers and vehicles heterogeneities play a major role in the oscillation process. For example, stop-and-go waves
may appear when several trucks with low acceleration rate are in the vicinity of the merge at the same time or try to
overtake each other. Investigations are made in this direction. They will take benefit of the general framework
proposed in (Laval and Leclercq, 2010b) to account for drivers’ aggressiveness.
Another great interest of the proposed analytical model is that it can improve any merge model based on the
demand / supply framework. We have shown in this paper how to implement it in the ND model by modifying the
supply function. This also applies to other classical macroscopic merge models, e.g. (Jin and Zhang, 2003; Lebacque
and Koshyaran, 2005) or to microscopic merge models that explicitly accounts for downstream supply (Chevallier
and Leclercq, 2009). Indeed, the analytical model directly assesses the capacity drop that reduces the available
downstream supply when the merge is active. Note that it can account for the length of the insertion section even if
the merge is represented as a point. This model can also be used to test any microscopic behavioral merge models at
a macroscopic scale. Indeed, it provides an estimation of the mean capacity drop that should be observed depending
on the different parameters. The authors are currently investigating the microscopic merge model proposed in (Laval
and Leclercq, 2008) to test its consistency at the macroscopic scale with the findings of this paper and its ability to
properly reproduce lane flow distributions and to account for vehicles heterogeneity. Preliminary results are
encouraging but research in this realm has to continue.

Acknowledgements

The authors are grateful to the Highways Agency of England for providing data on M6 freeways. The authors
would also greatly thank Cécile Becarie for her daily assistance and Estelle Chevallier for her careful reading of this
paper.

References

Ahn, S., Cassidy, M.J., 2007. Freeway Traffic Oscillations and Vehicle Lane-Change Maneuvers, In: Allsop, R., Bell, M.G.H., Heydecker, B.G.
(Eds.). Proceedings of the 17th International Symposium on Transportation and Traffic Theory, Londres: Elsevier, 691-710.
 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26 25

Aw, A. and Rascle, M., 2000. Resurrection of « second order » models of traffic flow ? SIAM Applied Mathematics, 60, 916-938.
Bar-Gera, H., Ahn, S., 2010. Empirical macroscopic evaluation of freeway merge-ratios, Transporation Research Part C, 18(4), 457-470.
Cassidy, M. J., Bertini, R. L., 1999. Some traffic features at freeway bottlenecks, Transportation Research Part B, 33(1), 25-42.
Cassidy, M.J., Ahn, S., 2005. Driver turn-taking behavior in congested freeway merges. Transportation Research Record - Journal of the
Transportation Research Board, 1934, 140-147.
Cassidy, M.J., Rudjanakanoknad, J., 2005. Increasing capacity of an isolated merge by metering its on-ram. Transportation Research Part B,
39(10), 896-913.
Chevallier, E., Leclercq, L., 2009. Do microscopic merging models reproduce the observed priority sharing ratio in congestion? Transportation
research Part C, 17(3), 328-336.
Chiabaut, N., Leclercq, L., Bretin, T., Buisson, C., 2008. Calibrating the Fundamental Diagram in Congestion: Methods Based on Observations at
Consecutive Loop-Detectors. Proceedings of the IWTDCS Conference, 8-9 September, Barcelona, (Spain) [CDROM]. Tokyo: University of
Tokyo, 9 p.
Chung, K., Rudjanakanoknad, J., Cassidy, M., 2007. Relation between traffic density and capacity drop at three freeway bottlenecks,
Transportation Research part B, 41(1), 82-95.
Daamen, W., Loot, L., Hoogendoorn, S. 2010. Empirical analysis of merging behavior at a freeway on-ramp. Proceedings of the Transportation
Research Board 89th Annual Meeting. Washington: TRB (DVD-Rom), 15 p.
Daganzo, C.F., 2005. A variational formulation of kinematic waves: basic theory and complex boundary conditions. Transportation Research part
B, 39(2), 187-196.
Daganzo, C., 1995. The cell transmission model, Part II: network traffic. Transportation Research part B, 29(2), 79-93.
Elefteriadou, L., Roess, R.P., McShane, W.R., 1995. Probabilistic nature of breakdown at freeway merge junctions, Transportation Research
Record - Journal of the Transportation Research Board, 1484, 80-89.
Evans, J., Elefteriadou, L., Natarajan, G., 2001. Determination of the probability of breakdown on a freeway based on zonal merging
probabilities. Transportation Research part B, 35(3), 237–254.
Greenberg, J., 2001. Extensions and amplifications of a traffic model of Aw and Rascle, SIAM Journal on Applied Mathematics, 62, 729–745.
Kerner, B.S., 2000. Theory of breakdown phenomenon at highway bottlenecks. Transportation Research Record - Journal of the Transportation
Research Board, 1710, 136–144.
Kerner, B.S., 2002. Empirical macroscopic features of spatial-temporal traffic patterns at highway bottlenecks, Physical Review E - Statistical,
Nonlinear, and Soft Matter Physics, 65(4), 1-30.
Kerner, B.S., 2004. The Physics of Traffic. Heidelberg: Springer, 296 p.
Koshi, M., Iwasaki, M., Okhura, I., 1983. Some findings and an overview on vehicular flow characteristics, In: Hurdle, V.F., Hauer, E., Steuart,
G.F. (Eds.) Proceedings of the 8th International Symposium on Transportation and Traffic Theory, Toronto: University of Toronto Press,
403-426.
Hall, F.L., Hall, L.M., 1990. Capacity and speed-flow analysis at the queen Elisabeth way in Ontario, Transportation Research Record - Journal
of the Transportation Research Board, 1287, 108-118.
Jin, W.L., Zhang, H.M., 2003. On the distribution schemes for determining flows through a merge. Transportation Research Part B, 37(6), 521-
540.
Laval, J.A., Cassidy, M.J., Daganzo, C.F., 2005. Impacts of lane changes at on-ramp bottlenecks: a theory and strategies to maximize capacity.
In: Kühne, R., Poeschl, T., Schadschneider, A., Schreckenberg, M., Wolf, D. (Eds.), Traffic and Granular Flow ‘05’. Berlin: Springer, 577-
586.
Laval, J.A., Daganzo C.F., 2006. Lane-changing in traffic streams. Transportation Research Part B, 40(3), 251-264.
Laval, J.A., Chen, D., Ben Amer, K., Guin, A., Ahn, S., 2009. Evolution of Oscillations in Congested Traffic: Improved Estimation Method and
Additional Empirical Evidence. Proceedings of the Transportation Research Board 88th Annual Meeting. Washington: TRB (DVD-Rom),
18 p.
Laval, J.A., Leclercq, L., 2008. Microscopic modeling of the relaxation phenomenon using a macroscopic lane-changing model. Transportation
Research part Part B, 42(6), 511-522.
Laval, J.A., Leclercq, L., 2010. Continuum Approximation for Congestion dynamics along freeway Corridors. Transportation Science, 44, 2010,
87-97.
Laval, J.A., Leclercq, L., 2010b. A mechanism to describe the formation and propagation of stop-and-go waves in congested freeway traffic.
Philosophical Transactions of Royal Society A, 368(1928), 4519-4541.
Lebacque, J.P., Lesort, J.B., Giorgi, F., 1998. Introducing buses into first-order macroscopic traffic flow models. Transportation Research
Record, 1644, 70-79.
Lebacque, J.P., Koshyaran, M.M., 2005. First order macroscopic traffic flow models: intersections modeling, network modeling. In: Mahmassani,
H.S. (Ed.), 16th ISTTT, Pergamon, London, 365-386.
Leclercq, L., Chanut, S., Lesort, J.B., 2004. Moving bottlenecks in the LWR model: a unified theory. Transportation Research Record, 1883, 3-
13.
Leclercq, L., 2005. A new numerical scheme for bounding acceleration in the LWR model. In: Heydecker, B. (Ed.) 4th IMA Conference on
Mathematics in Transport studies, 7-9 Septembre, London, 279-292.
Li, X., Peng, F., Ouyang, Y., 2010. Measurement and estimation of traffic oscillation properties, Transportation Research Part B, 44(1), 1-14.
Lighthill, M.J and Whitham, J.B., 1955. On kinematic waves II: A theory of traffic flow in long crowded roads. Proceedings of the Royal
Society, A229, 317-345.
Mauch, M., Cassidy, M. J., 2002. Freeway traffic oscillations: observations and predictions. In: Taylor, M.A.P (Ed.) Proceedings of the 15th
International Symposium on Transportation and Traffic Theory, Amsterdam : Pergamon, 653-674.
26 Ludovic Leclercq et al. / Procedia Social and Behavioral Sciences 17 (2011) 12–26

Munoz, J.C., Daganzo, C.F., 2002. Moving Bottlenecks: A Theory Grounded on Experimetnal Observation. In: Taylor, M. (Ed.) 15th ISTTT,
Elsevier, London, 441-462.
Newell, G.F., 1982. Applications of queueing theory (2nd edition). New York: Chapman & Hall, 303 p.
Newell, G.F., 1998. A moving bottleneck. Transportation Research Part B. 32(8), 531-537.
Oehlert, G. W., 1992. A note on the Delta method. The American Statistician 46(1), 27-29.
Persaud, B., Yagar, S., Brownlee, R., 1998. Exploration of the Breakdown Phenomenon in Freeway Traffic. Transportation Research Record -
Journal of the Transportation Research Board, 1634, 64-69.
Richards, P.I., 1956. Shockwaves on the highway. Operations Research, 4, pp. 42-51.
Sarvi, M., Kuwahara, M., Ceder, A., 2007. Observing freeway ramp merging phenomena in congested traffic, Journal of Advanced
Transportation, 41(2), 145-170.
Siebel, F., Mauser, W., Moutari, S., Rascle, M., 2009. Balanced vehicular traffic at bottleneck, Mathematical and Computer Modelling, 49(3-4),
689-702.
Siebel, F., Mauser, W., Moutari, S., Rascle, M., 2009b. Modeling synchronized flow at highway bottlenecks, In: Appert-Rolland, C., Chevoir, F.,
Gondret, P., Lassarre, S., Lebacque, J.P., Schreckenberg, M. (Eds), Traffic and Granular Flow ‘07’. Berlin: Springer, 201-210.
Treiber, M., Kesting, A., Helbing, D., 2006. Understanding widely scattered traffic flows, the capacity drop, and platoons as effects of variance-
driven time gaps. Physical Review E - Statistical, Nonlinear, and Soft Matter Physics, 74(2), 1-10.
Troutbeck, R.J., 2002. The performance of uncontrolled merges using a limited priority process. In: Taylor, M.A.P., (Ed.), Proceedings of the
15th International Symposium on Transportation and Traffic Theory, Amsterdam : Pergamon, 463-482.
Yi, H., Mulinazzi, T., 2007. Observed distribution patterns of on-ramp merge lengths on urban freeways, Transportation Research Record -
Journal of the Transportation Research Board, 2023, 120-129.