Transportation Research Part A 80 (2015) 320–329

Contents lists available at ScienceDirect

Transportation Research Part A

journal homepage: www.elsevier.com/locate/tra

A problem of limited-access special lanes. Part II: Exploring

remedies via simulation
Michael J. Cassidy a, Kwangho Kim b, Wei Ni a, Weihua Gu c,⇑
a
Department of Civil and Environmental Engineering and the Institute of Transportation Studies, University of California, Berkeley, CA 94720, USA
b
National Infrastructure Research Division, Korea Research Institute for Human Settlements, Anyang, Republic of Korea
c
Department of Electrical Engineering, The Hong Kong Polytechnic University, Hung Hom, Hong Kong

a r t i c l e i n f o a b s t r a c t

Article history: Spatiotemporal analyses of freeway sites in Part I have shown that special-lane access
Available online 29 July 2015 points are prone to become bottlenecks. These can degrade traffic flows, sometimes in
all lanes. Part II explores select impacts of re-designing the means of entering and exiting
Keywords: a special lane, and of altering the policy governing its use. Parametric tests were conducted
Managed lanes using a computer simulation model that was calibrated to one of the sites studied in Part I;
Carpool lanes one with a buffer-separated carpool lane. Though less reliable than what might have been
Toll lanes
observed via experiments in real settings, the simulated findings seem to offer useful
Traffic simulation
insights nonetheless.
The findings indicate that traffic conditions would improve at the site by elongating the
carpool lane’s buffer opening beyond its present length of 400 m. Yet, only modest
improvements were predicted, even when the opening was elongated to 1000 m or more.
Greater benefits were predicted from disentangling the movements made into and out of
the carpool lane. This was achieved by placing first a buffer opening to serve only ingress,
followed by another immediately downstream to serve egress. The benefits of this treat-
ment were again limited, even when each tandem opening was elongated to a length of
700 m. Fully removing the buffer that physically separates the carpool lane from the reg-
ular ones was predicted to bring the greatest improvements to traffic. Also examined
was pending legislation that would leave the carpool-lane buffer in place, while limiting
the times of day when the lane is reserved for special use. Simulations predict that this leg-
islation would degrade travel conditions below those that presently occur at the site. The
extent of this predicted degradation varied, depending upon the time of day when the
lane-use restriction went into effect.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction

Part I established that the access points to special-use freeway lanes can become bottlenecks. Part II pursues a handful of
possible solutions. Exploring solutions by means of experiments conducted in real settings would have required cooperative
agreements with highway departments that are difficult to obtain. Moreover, the attendant costs of field experiments would
have severely limited their scope.

⇑ Corresponding author. Tel.: +852 2766 6189; fax: +852 2330 1544.
E-mail address: weihua.gu@polyu.edu.hk (W. Gu).

http://dx.doi.org/10.1016/j.tra.2015.07.003
0965-8564/Ó 2015 Elsevier Ltd. All rights reserved.
 M.J. Cassidy et al. / Transportation Research Part A 80 (2015) 320–329 321

Computer simulation was therefore used instead. A well-tested model of driver car-following and lane-changing behavior
was chosen for this purpose. The model was calibrated to match observations from one of the freeway sites studied in Part I;
one that features a limited-access carpool lane with a 400 m-long buffer opening. Details on these matters are offered in the
following section. Parametric tests using the calibrated model are described thereafter.
The first round of parametric tests is presented in Section 3. It features experiments in which access to the site’s carpool
lane was redesigned, first by (i) elongating the buffer opening incrementally; then by (ii) inserting openings of various
lengths in tandem, with one opening for ingress and the other for egress; and finally by (iii) fully removing the buffer, such
that the carpool lane became a non-separated one. Benefits were predicted under all three treatments, though not in equal
measure. Traffic generally benefited more from tandem openings, as compared against an elongated single one. Yet neither
of these first two treatments completely eliminated the flow disruptions caused by focused ingress and egress, even when
tested using travel demands that were relatively small. Greatest benefits were predicted when the buffer was fully removed
from the site, despite conservative assumptions made in regard to how carpoolers might respond to that removal. Yet even
this treatment was not a silver bullet against congestion: buffer removal eased the constraints on merging and diverging at
nearby ramp junctions, and a new bottleneck triggered by weaving maneuvers was predicted to emerge as a result. This new
bottleneck was a relatively minor one, however, and was less damaging to flows.
Section 4 examines pending legislation under which the site’s limited-access lane would be reserved for carpools only
during the rush, and made available to all traffic at all other times of the day. Heightened vehicle migration through the
400 m-long buffer opening was predicted whenever the lane switched from general to restricted use, which disrupted flows.
The extent of the disruption varied, depending upon the times of day when the carpool restriction took effect. Interestingly,
the parametric tests indicate that, all else equal, a time-of-day restriction is always more damaging to flow at the site than is
the current policy in which carpool-lane use is restricted at all times.
Section 5 discusses policy implications of the present findings. Future steps for moving forward are discussed as well.

2. Computer simulation

All experiments were conducted using a discrete-time, microscopic car-following and lane-changing model formulated in
Lee (2008). The model was developed expressly for weaving sections, where a freeway’s geometry causes certain vehicle
streams to cross one another (e.g. see TRB, 2010), much as they do at special-lane access points. The model in Lee (2008)
is itself an adaptation to one presented in Menendez and Daganzo (2007).
To simulate car-following, vehicles are assumed at each time step to have traveled the greatest distances possible, subject
to constraints imposed by: maximum vehicle accelerations, driver comfort, and safety. Full descriptions of the car-following
algorithms are furnished in Menendez (2006).
Vehicular lane-changing maneuvers are classified as being either (i) ‘‘mandatory,’’ which in the present context entail
movements into and out of the carpool lane, as well as merge and diverge maneuvers at on- and off-ramps; or (ii) ‘‘optional,’’
which entail the maneuvers that are performed between regular lanes as drivers unilaterally seek to increase their own tra-
vel speeds.
Optional maneuvers were modeled as per the logic of Menendez and Daganzo (2007), and the reviewer can refer to that
reference for details. Mandatory maneuvers were emulated using the logit model of Lee (2008). That model takes as input:
the difference in traffic densities between a driver’s present lane and her adjacent target lane; the number of lanes that a
vehicle must cross to complete the maneuver; and the longitudinal road space available to do so. When a vehicle fails to
perform a mandatory maneuver over a sufficient number of time steps, that vehicle or its neighbor in the target lane will
decelerate to increase the likelihood of a successful lane change; see Lee (2008) for further details.
Review of Lee (2008) shows that the model fared well when compared against real data from freeway weaving sections. It
is therefore not surprising that the model was able to match observed conditions at an access-point bottleneck fairly well, as
shown below.

2.1. Model calibration

The study site was taken to be the stretch of Interstate 210 previously shown in Fig. 3a of Part I, which features a
limited-access carpool lane with a painted buffer and a 400 m access point to serve ingress and egress. The site is reproduced
in Fig. 1 below for the reader’s convenience. The simulation model was calibrated to the 1490 m portion of the site that spans
the detectors at D2 and D1, and to the traffic conditions measured from 14:30 h to 16:00 h on March 14, 2012. That day was
selected because it was largely free of flow disruptions caused by queues spilling-over from somewhere downstream of the
site. Simulations could therefore be performed without the added complication of having to estimate travel demands during
periods when flows were constrained by spill-over queues.
Similar complications nonetheless occurred when the queue that formed that day at the access point engulfed the site’s
upstream portion. Once that queue had propagated beyond the detectors at D2, the flows measured there were constrained
and were therefore lower than the demand. We were thus forced to calibrate the model by guessing at the demand to pass
D2 during that constrained period. We guessed this to be the highest 5-min flow measured at D2 prior to the queue’s arrival,
322 M.J. Cassidy et al. / Transportation Research Part A 80 (2015) 320–329

Fig. 1. Eastbound I-210 Freeway, Los Angeles County, California.

8630 vph. We think this to be a low-end estimate. For the parametric analyses to come in Sections 3 and 4, carpool-lane
designs and policies governing use will be explored for that low-end estimate, and for higher demand estimates as well.
It was also necessary to estimate time-varying demands disaggregated by four movements, these being the demands: to
traverse the stretch from D2 to D1 while remaining in regular lanes; to traverse the stretch while remaining in the carpool
lane; and to enter and exit the carpool lane via the site’s 400 m-long access point. These, along with the time-varying
demands for the site’s nearest downstream off-ramps, were estimated in simple ways that are described in Appendix A.
Merge flows from the site’s metered on-ramps were measured from data.
Numeric values for model parameters such as free-flow vehicle speed, vehicle acceleration, jam spacing and queued wave
speed were taken directly from Menendez (2006) and Lee (2008). Other parameter values were calibrated to the measured
data. Parameter values, including those selected via the calibration process, are presented in Table 1. Outcomes of the cal-
ibration process are presented below.
We start with qualitative comparisons between simulated congestion patterns on the site and the real (i.e. measured)
ones. Note from visual inspection of Fig. 2a and b how the carpool lane’s simulated time–space–occupancy diagram matches
its measured counterpart rather well. The same can be said for the simulated and measured patterns in adjacent regular-lane
2; see Fig. 3a and b.
For a more quantitative assessment, we take the metric to be the time-varying vehicle accumulations in all lanes between
D2 and D1, because accumulations are notoriously difficult to predict with simulation; e.g. see Daganzo (1994). Fig. 4 presents
the measured 5-min average accumulations (bold curve) along with the simulated values (thin curve). In most instances, the
difference between measured and simulated values is less than 5%, and even the largest difference is less than 9%.
Though we are unaware of any standardized target for matching measured and simulated accumulations, we interpret
agreements to within 10% to be quite good, and therefore use the calibrated model in the parametric tests to follow. The
resulting predictions are surely not precise depictions of reality. Still, we believe them to be useful approximations, partic-
ularly for making coarse, macro-level comparisons, whether comparing outputs across distinct inputs for a given
carpool-lane design or policy, or across distinct designs and policies.

3. Design alterations

Impacts of incrementally elongating the access point under varying rush-hour travel demands are studied in the follow-
ing subsection. Limited-access designs with separate buffer openings for ingress and egress are thereafter examined in

Table 1
Model parameters.

Parameter Description Value

Parameters of traffic and vehicle kinetics
u Vehicle free-flow speed 31.1 m/s (110 km/h)
aU Maximum vehicle acceleration rate 3.0 m/s2
r Dimensionless proxy for maximum deceleration rate 0.9
x Backward wave speed 6.3 m/s
Sjammed Jammed vehicle spacing 6.1 m
Parameters in the logit model for mandatory lane changes
b0 Constant 8.0
b1 Coefficient of density difference between the original and target lanes 0.9
b2 Coefficient of inverse of normalized distance to the end of the access point opening 6.3
b3 Coefficient of number of lanes to be crossed 1.2
Other parameters
w Cooperation initiation time for mandatory lane changes 5.0 s
/ Sensitivity to relative differences in speed between adjacent lanes for optional lane changes 4.0 s
 M.J. Cassidy et al. / Transportation Research Part A 80 (2015) 320–329 323

Fig. 2. Time–space–occupancy plots of carpool lane (March 14, 2012). (a) Simulated. (b) Measured.

similar parametric fashion. The section ends with experiments in which the painted buffer between carpool and regular lane
is removed, and carpoolers respond (conservatively) by spatially-spreading their ingress and egress maneuvers by modest
amounts.
Simulated experiments were performed for the 90-min period from 14:30 to 16:00 h. The first 30 min, from 14:30 to
15:00, served as a warm-up period. The time-varying demands to pass detector station D2 during that period were the rates
actually measured at D2. Simulated outcomes were extracted only from the final 1-h period from 15:00 to 16:00. In this way,
the outcomes reflect conditions occurring during the rush, and are by-and-large unaffected by the lower, pre-rush demands.
Rush-hour demands to pass upstream detector station D2 were examined in parametric fashion. These ranged from
8630 vph, the estimate used in Section 2 for model calibration, to 9920 vph, a 15% increase. The two other estimated
demands examined in these tests reflect 5% and 10% increases over the base estimate of 8630 vph.
Each outcome to be presented below is the vehicle speed predicted for the site, as averaged: over the 1-h rush; across all
lanes, including the carpool lane; over the site’s full 1490 m length; and over 3 separate simulations with inputs held fixed.
Average speed was chosen as the performance metric because a predicted average that falls below the designated free-flow
speed (110 km/h, see again Table 1) can be taken as an unambiguous indication of congestion; and because comparing the
congested speeds predicted under distinct carpool-lane designs is a simple and straightforward way to assess the relative
merits of those designs.1

3.1. Elongating the access point

Fig. 5 presents the average speeds predicted for access-point lengths ranging from the existing length of 400 m, to
1200 m. Each curve in the figure presents the predicted speeds for a distinct rush-hour demand to pass D2. The figure shows
that congestion persists under all scenarios examined; i.e. note that even when demand is low and the access point is long,
average speed falls below 70 km/h, which is well below the designated free-flow speed of 110 km/h that characterizes
uncongested traffic.
Fig. 5 further reveals how congestion tends to worsen (i.e. average speeds typically diminish) as demand increases. This
occurs, in part, because demands to maneuver through the access point increase as demand to pass D2 increases. The effects
(i.e. the vertical displacements between curves) are rather small when the access points are short in length. This is because
higher ingress and egress demands eroded conditions on the site only in the earliest portions of the rush, before flows enter-
ing the site became constrained by queues.
As the access points grew longer, the bottleneck problems diminished. This is why average speeds on the site tended to
increase with longer access points, as is clear from the curves. This also meant that the higher ingress and egress demands

1
Average speed is a similarly suitable metric for comparing across policies governing carpool-lane use. That metric shall therefore be used in Section 4 as
well.
324 M.J. Cassidy et al. / Transportation Research Part A 80 (2015) 320–329

Fig. 3. Time–space–occupancy plots of regular-lane 2 (March 14, 2012). (a) Simulated. (b) Measured.

Fig. 4. Time-series of 5-min vehicle accumulations across all travel lanes (I-210E; March 14, 2012).

Fig. 5. Average speeds under varying access-point lengths and rush-hour demands.
 M.J. Cassidy et al. / Transportation Research Part A 80 (2015) 320–329 325

Fig. 6. Average speeds under varying lengths of tandem openings and varying rush-hour demands.

exerted their influence over longer portions of the early rush. This is why the vertical displacements between curves increase
as access points grow in length.
Of further interest, all the curves in Fig. 5 display concave shapes, indicating that gains in speed gradually diminish with
each incremental elongation of the access point. And the gains approach zero once the buffer opening exceeds 1000 m.
The pattern is a consequence of a realistic feature of the logit model of Lee (2008). The model emulates a certain observed
behavior of drivers when traversing freeway weaving sections, namely: drivers who perform (mandatory) weaving maneu-
vers tend to do so soon after entering a weaving section. Few drivers defer maneuvers until ‘‘late in the game’’ when little
longitudinal space remains available for executing them. We observed similar sorts of ingress and egress tendencies at the
study site.2 Thus for carpool-lane access points of 800 m or longer, the model predicts that roughly 80% of all ingress and egress
maneuvers are performed within the first 750 m of those openings. In short, it appears that elongating the site’s access point
beyond 800 m or so would offer little benefit in terms of diffusing ingress and egress, and of diminishing the attendant disrup-
tions to flow.

3.2. Disentangling ingress from egress

Next examined are the conditions predicted when the site’s limited-access carpool lane featured two closely-spaced
buffer openings in tandem. The upstream opening was reserved for ingress; the downstream one for egress; and the two
openings were separated by 200 m of solid-painted lane line.3
Parametric analysis was used to explore how the length of each buffer opening affects vehicle speed at the site. Each sce-
nario tested featured tandem openings of equal lengths, so as to limit dimensionality. These openings ranged in length from
400 m to 700 m. Though not presented in the interest of brevity, simulations predicted that openings less than 400 m in
length would be very disruptive for the range of rush-period demands examined, primarily because ingress vehicles con-
fronted with short openings had difficulty finding suitable gaps in carpool-lane traffic.
Outcomes are presented in Fig. 6. Separate curves again denote distinct rush-hour demands to pass D2, and these are the
same demands used in Section 3.1. As expected, speeds increase as the tandem openings expand. Yet much like in
Section 3.1, speed gains tend to diminish gradually with each elongation of the openings. Much as before, the pattern is
caused by drivers’ propensity to maneuver sooner, rather than later, upon reaching a buffer opening. The curves indicate that
elongating the tandem openings beyond about 600 m each would bring little added benefit to traffic flows on the study site.
Of further note, tandem designs compare favorably to conventional single-access-point designs. Visual comparisons of
Figs. 5 and 6 reveal that tandem openings of only 400 m each produce average vehicle speeds comparable to what are pre-
dicted for a single, elongated access point of 1200 m length. And tandem designs steadily become even better options as the
ingress and egress openings are expanded. Note further from Figs. 5 and 6 that once the tandem openings each reach 700 m,
the resulting average speeds are roughly 10 km/h higher than they would be in the presence of a single 1200 m-long buffer
opening. And note from those same figures that this was predicted to be the case for all demands examined.
Yet tandem designs do not hold-up well when compared against what is predicted to occur when the carpool lane’s
painted buffer is fully removed. This is explored as part of the simulations to follow.

2
Recall from Part I that traffic on the site was videotaped during an afternoon in 2013. The aforementioned driver tendency was evident in the video images.
3
Simulations predicted that swapping this sequence (such that an opening for egress preceded one for ingress) would produce disruptive flow patterns that
degrade travel in all lanes. This state of affairs arose primarily because simulated vehicles that egressed from the carpool lane via its upstream opening tended
thereafter to linger just long enough in the adjacent (left-most) regular lanes so as to impede other vehicles seeking ingress downstream. The resulting damage
was predicted to be so severe that we do not report further on this ill-fated design option.
326 M.J. Cassidy et al. / Transportation Research Part A 80 (2015) 320–329

3.3. Removing the buffer

Using the select model to simulate physically unrestrained carpool-lane ingress and egress was easier said than done.
Removing the painted buffer would distribute carpoolers’ in and out movements over greater space.4 Yet, the precise pattern
of this redistribution is unknown. We therefore have no basis upon which to re-calibrate the logit model of Lee (2008). In light of
this uncertainty, the redistribution of ingress and egress was coarsely modeled in parametric fashion. Conservative assumptions
were adopted, to avoid over-stating the benefits of buffer removal. The experiments were conducted in the following fashion.
The same simulation model, with its attendant logit model, was used in the experiments, despite the buffer’s removal
from the scene. Staying with the same model ensured that predicted outcomes could be compared commensurately against
those involving the other (i.e. limited-access) carpool-lane designs of Sections 3.1 and 3.2, as well against the distinct
lane-use policies to be examined in Section 4. On the downside, the model, when left to its own devices, seemingly did
not do a good job of emulating the spreading of ingress and egress induced by buffer removal. It predicted instead that most
of those maneuvers took place over the 750 m stretch that extends forward from the site’s on-ramps (see again Fig. 1) due to
the attribute of the attendant logit model, as already explained in Section 3.1.
It was therefore assumed that, in the absence of the buffer, portions of the ingress and egress movements now occurred
outside the bounds of the 1490 m-long study site. This would be the case if some carpoolers deferred those maneuvers to
locations downstream of detector station D1, for example; and these deferments served as means to approximate the spatial
redistributions of interest. To be conservative, the deferment amounts were kept small: they were varied over a series of
experiments from 0% to a paltry 10% of the ingress and egress numbers estimated from the real data.
For added realism, and to stay conservative, the vehicle speeds predicted for the carpool lane were systematically
diminished so as to emulate the frictional effect, whereby queued, slow-moving traffic in regular lanes slowed traffic in
the adjacent (non-separated) carpool lane as well; see Jang and Cassidy (2012). Fig. 5 of that aforementioned reference
was used to estimate speed-reduction factors due to that effect. The selected factors diminished carpool-lane speeds by
about 9%.
Outcomes of the simulations are presented in Fig. 7a. Its curves display the average rush-hour speeds as functions of the
percent of ingress and egress deferred. Separate curves once again distinguish our estimated demands to pass D2.
The curves’ slopes reveal that the gains in average speed were relatively large as the percent deferment grew from 0% to
5%. This illustrates (approximately) how the spatial spreading of carpool-lane ingress and egress can lessen flow disruptions.
The greatest gains in speed occurred in the median lanes, as one might expect, given that these lie adjacent to the erstwhile
access point and were therefore subject to the greatest flow reductions due to focused ingress and egress. As an example, the
bold curve in Fig. 7b shows how speeds rose in regular (i.e. near-median) lane 2 as deferment increased, and for our lowest
estimate of rush-hour demand to pass D2, 8630 vph.
Returning to Fig. 7a, note how its curves also reveal that speed gains gradually diminished as the percent deferment
approached 10%. It turns out that as deferments grew, the damaging disruptions formerly brought by focused ingress and
egress through the access point were gradually supplanted by a new cause of flow disruption, that being: merging and
diverging maneuvers near the site’s on- and off-ramps. As movements into and out of the carpool lane were gradually
deferred in the buffer’s absence and thus became less damaging, constraints were lessened on the merge and diverge flows
near the ramps. The outpouring of these latter flows became damaging in their own rights; i.e. they created a so-called weav-
ing bottleneck on the site.
This weaving bottleneck was most damaging to traffic in the shoulder lane, which is not surprising given its adjacency
to the ramps. For illustration, the thin curve in Fig. 7b shows how shoulder-lane speeds diminished as deferment
increased.
So it seems that removing the access-point bottleneck would not be a ‘‘silver bullet’’ to prevent congestion at the site.
Yet the weaving bottleneck predicted to supplant the site’s existing one would be less damaging to flows, as evident in
Fig. 7c. The curve in that figure was constructed from select outcomes previously shown in Figs. 5 and 7a. It reveals
the gains in speed brought by buffer removal, DS, as a function of percent deferment. The DS were computed using as
benchmarks: the low estimate of demand to pass D2, 8630 vph; and the access-point length of 400 m that currently exists
at the site.
The reader can verify that use of other benchmark demands and access-point lengths yield similar curves of DS. Hence,
we see that buffer removal would be a boon to vehicle speed on the site, as compared against limited-access designs that
feature a single buffer opening.5 Note from Fig. 7c how this is predicted to be the case, even if buffer removal were to create
little or no deferment in carpool-lane ingress and egress.
Moreover, the speeds predicted once the buffer was removed are typically higher than those predicted for limited-access
designs that feature two buffer openings in tandem. This becomes evident from visual comparisons of Figs. 6 and 7a. Note
from these comparisons that if buffer removal can generate even a paltry 2.5% deferment in ingress and egress, then it wins
over a tandem design, even when the latter features highly elongated openings of 700 m each. Should buffer removal gen-
erate deferments that exceed 2.5%, then it becomes the superior option by a wider margin still.

4
We know this because non-separated carpool lanes are not prone to become bottlenecks; see Cassidy et al. (2010).
5
This was predicted to be the case even for carpool-lane vehicles, which would suffer the frictional effect. After accounting for this effect, carpool-lane speeds
still increased by more than 8%.
 M.J. Cassidy et al. / Transportation Research Part A 80 (2015) 320–329 327

Fig. 7a. Average speeds under buffer removal (with varying demands and varying deferred ingress and egress).

Fig. 7b. Average speeds under buffer removal for shoulder lane and lane 2 (demand to pass D2 = 8630 vph).

Fig. 7c. Gains in average speed under buffer removal with varying deferred ingress and egress (demand to pass D2 = 8630 vph).

4. Time-of-day restrictions

The final set of experiments examines the impacts of pending legislation that would limit the carpool restriction presently
imposed on the study site, as well as on select other freeway sites in southern California; see California Legislative
Information (2013). We believe that this examination could be of interest to highway departments both inside and outside
California, since any number of agencies may be contemplating similar legislation.
328 M.J. Cassidy et al. / Transportation Research Part A 80 (2015) 320–329

Table 2
Tests of lane-use policies.

Lane-use policy Average speed from 14:00 to 16:00 (km/h)

Demand to pass D2 = 8630 vph Demand to pass D2 = 9920 vph
Day-long restriction 66 63
Time-of-day restriction starting at 14:30 h 61 59
Time-of-day restriction starting at 15:00 h 56 52

Under the proposed law, select carpool lanes in southern California would evidently remain limited-access ones. Yet,
restrictions on the use of those lanes would be eased: the lanes would be reserved for carpools during weekday rush hours
(only), and be made available for general use at all other times.
To predict the impacts of this law, the experiments featured a simulation period that began 30 min earlier than in our
previous tests, and thus extended from 14:00 to 16:00 h. The longer period was used to accommodate a 14:30 start time
for the lane-use restriction. The time 14:30 is 30 min earlier than the start time uniformly instituted at freeway carpool facil-
ities in California where time-of-day restrictions are in place. Yet, this earlier time turns out to have an apparent advantage:
the heightened ingress and egress that would accompany the activation of the lane-use restriction would occur prior to the
emergence of the site’s persistent rush-hour queues. The activation was therefore predicted to be less damaging to traffic, as
will be shown momentarily. California’s customary start time for carpool restrictions, 15:00 h, was separately tested as well.
In all instances, drivers were programmed to respond to an impending time-of-day restriction several minutes earlier than
its officially scheduled start. Prior to those responses, drivers were assumed to distribute themselves evenly over all freeway
lanes. Speeds averaged across all lanes were computed for the entire 2-h period from 14:00 to 16:00. The buffer opening was
left at 400 m, which is presumably what would occur at the study site, should the pending legislation come to pass. All else
regarding the following experiments were as previously described in Section 3.
Interestingly, the legislation represents an attempt to improve travel conditions by filling-up carpool lanes during
off-peak periods, when the lanes tend otherwise to be under-used; see again California Legislative Information (2013).
However, the simulations predict that the law would backfire, at least for the case of our study site. Outcomes of the exper-
iments are summarized in Table 2, both for the lowest- and highest estimate of rush-hour demand to pass D2.
The table shows that activating the lane-use restriction at 14:30 would produce lower average speeds than does the pol-
icy of day-long restriction currently in force.6 As previously noted, this negative finding occurs due to the heightened ingress
and egress that occurred near the 14:30 start time. These heightened maneuvers disrupted and slowed traffic. Because the dis-
ruptions occurred early enough in the afternoon, the resulting queues vanished prior to the emergence of persistent,
rush-period congestion at around 15:00 h. But the damage was done; i.e. the early queues at around 14:30 – though transient
– pulled down the 2 h-average of the site’s vehicle speeds, as is evident in the table.
Activating the lane-use restriction at 15:00 h, as is the custom in California, would be more damaging still. In this latter
case, the heightened ingress and egress brought by the later-occurring activation would roughly coincide with the start of
persistent, rush-period queueing. Simulations predict that this would exacerbate queueing and diminish speeds substan-
tially, as compared against the simulated outcomes under the present, day-long-restriction policy. Table 2 shows that the
predicted reductions in average speed reach or exceed 10 km/h, both for the low- and high estimates of rush-hour travel
demand.

5. Conclusions

Simulations predict that an access-point bottleneck can be made less damaging to traffic by elongating the buffer open-
ing, or better still, by deploying tandem openings to disentangle ingress from egress. Still greater improvements were usually
predicted to occur by removing the carpool lane’s painted buffer, so that carpoolers were no longer constrained to maneuver
through some prescribed opening. The simulations predicted that even this latter intervention would not completely elim-
inate congestion from the site. Yet, the new bottleneck triggered by the buffer’s removal was predicted to be less damaging
than is the access-point bottleneck that currently resides at the site.
Pending legislation to limit the carpool-lane restriction to rush periods (only), while leaving the lane’s buffer in place,
would, according to the simulations, be a bad idea, at least during the rush period for the site studied here. The policy would
apparently be especially damaging if the lane-use restriction were to take effect at the start time customarily adopted in
California, since that time coincides with the onset of persistent rush-period queues at the site.
There is ample evidence that a time-of-day restriction, when imposed on a non-separated special-use lane, can be a very
good thing for freeway traffic; see again Part I of the present paper and see especially Cassidy et al. (2010). Problems are
predicted to occur only if the pending law were to seek a ‘‘short-cut’’ solution that changes the site’s lane-use policy, without
also physically changing the manner in which the carpool lane is to be accessed.

6
The site’s present-day conditions are simulated rather than estimated from real measurements in part so that the influence of travel demand could be
examined in parametric fashion.
 M.J. Cassidy et al. / Transportation Research Part A 80 (2015) 320–329 329

The present findings are not intended to serve as final word on what constitutes suitable designs or policies for
special-use freeway lanes. After all, the simulated predictions upon which Part II of this paper relies are imperfect, as sim-
ulated predictions always are. Moreover, the predictions in Part II are made only for a single site, and only for a limited num-
ber of possible remedies. For example, the work has not explored limited-access lane designs that feature ramp-type access
and egress of the kind in Goodin et al. (2013). Nor were issues of traffic safety and ease of enforcement presently explored.
Though these are important matters, they lie beyond what can be described realistically by our simulation model.
The above notwithstanding, it is hoped instead that the present findings – both in Parts I and II – might motivate further
thinking on how best to manage special-use lanes. A next logical step might be to pursue field experiments, to test
special-lane designs and policies in real settings. If done carefully, experiments of this kind might furnish findings that
are more definitive than what are presently offered in Part II. But success in this realm would require that the traffic impacts
of any and all designs and policies be field-measured over both time and space. Part I of this paper, as well as previous stud-
ies (e.g. see Cassidy et al., 2009), reveal that traffic measurements made in fragmented, purely temporal fashion can obscure
even sizable impacts of special lanes. This can lead to questionable conclusions regarding the efficacies of their designs and
of the policies that govern their use.

Appendix A. Estimating disaggregate demands

The appendix describes how demands for key traffic movements were estimated in simple ways from more aggregate
counts of traffic.
We estimate the demands at time, t, for trips between detector stations D2 and D1 that:

 remain in the carpool lane over the full trip, qCC(t);

 remain in regular lanes, qRR(t);
 exit the carpool lane via the site’s access point, qCR(t); and
 enter the carpool lane via that access point, qRC(t).

Prior to queues reaching the detector stations at D2, these disaggregate demands were equivalent to their corresponding
flows. These could be estimated from detector measurements at D2 and D1 using the following logic.

(1) The time-varying flows measured in the carpool lane at D2 were known to be qCC(t) + qCR(t).
(2) All of the qCR(t) were assumed to exit via the three off-ramps that reside between the site’s carpool-lane access point
and the next access point several miles downstream. Hence, this exiting carpool flow, qCR(t) = b(qCC(t) + qCR(t)), where
b is the fraction of carpoolers that exit via the three ramps. This exit fraction was assumed to be the same for carpool-
ers and non-carpoolers alike, and to hold both before and after rush-period queues engulfed detector station D2. From
flows measured on the freeway and the three off-ramps, the time-averaged b was estimated to be 0.32.
(3) The flows measured in the regular lanes at D2 along with those measured at the site’s on-ramps were the
qRR(t) + qRC(t).
(4) The differences in carpool-lane flows measured at D2 and D1 were the qRC(t) qCR(t), where measured differences at
D2 and D1 were synchronized to account for time-varying vehicle trip times between those two detector stations.

The above logic results in an easily-solvable system of 4 equations and 4 unknowns. The disaggregate flows were esti-
mated for 1-min time steps.
When queues eventually constrained flows at D2, the disaggregate demands were estimated from a presumed aggregate
demand to pass D2. Sections 3 and 4 specify these aggregate demands used in our parametric analyses.

References

California Legislative Information, 2013. Assembly Bill No. 405. <http://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=

201320140AB201320405>.
Cassidy, M.J., Daganzo, C.F., Jang, K., Chung, K., 2009. Spatiotemporal effects of segregating different vehicle classes on separate lanes. In: Lam, W., Wong,
S.C., Hong, K.L. (Eds.), Proceedings of the 13th International Symposium on Transportation and Traffic Theory, pp. 57–74.
Cassidy, M.J., Jang, K., Daganzo, C.F., 2010. The smoothing effect of carpool lanes on freeway bottlenecks. Transport. Res. Part A 44, 65–75.
Daganzo, C.F., 1994. The cell transmission model: a dynamic representation of highway traffic consistent with the hydrodynamic theory. Transport. Res.
Part B 28, 269–287.
Goodin, G., Benz, R., Burris, M., Brewer, M., Wood, N., Geiselbrecht, T., 2013. Katy Freeway: An Evaluation of a Second-generation Managed Lanes Project.
FHWA/TX-13/0-6688-1, Texas Department of Transportation and the Federal Highway Administration.
Jang, K., Cassidy, M.J., 2012. Dual influences on vehicle speed in special-use lanes and critique of US regulation. Transport. Res. Part A 46, 1108–1123.
Lee, J.H., 2008. Observations on Traffic Behavior in Freeway Weaving Bottlenecks: Empirical Study and Theoretical Modeling. Department of Civil and
Environmental Engineering, University of California, Berkeley.
Menendez, M., 2006. An Analysis of HOV Lanes: Their Impact on Traffic. Department of Civil and Environmental Engineering, University of California,
Berkeley.
Menendez, M., Daganzo, C.F., 2007. Effects of HOV lanes on freeway bottlenecks. Transport. Res. Part B 41, 809–822.
Transportation Research Board (TRB), 2010. Highway Capacity Manual. Transportation Research Board of the National Academies, Washington, D.C..