Transportation Research Part B 38 (2004) 441458

www.elsevier.com/locate/trb

Dynamic one-way trac control in automated

transportation systems
Mark Ebben

a,*

, Durk-Jouke van der Zee b, Matthieu van der Heijden

School of Business, Public Administration and Technology, University of Twente, P.O. Box 217,
7500 AE Enschede, The Netherlands
Faculty of Management and Organisation, University of Groningen, P.O. Box 800, 9700 AV Groningen,
The Netherlands
Received 7 September 2000; accepted 23 May 2003

Abstract
In a project on underground freight transportation using Automated Guided Vehicles, single lanes for
trac in two directions are constructed to reduce infrastructure investment. Intelligent control rules are
required to manage vehicle ows such, that collision is avoided and waiting times are minimised. In contrast
to standard trac control at intersections, these control rules should take into account signicant driving
times along the single lane (in our application up to 8 min). Whereas periodic control rules are often applied
in trac theory, we focus on adaptive rules such as look-ahead heuristics and dynamic programming algorithms. Numerical experiments show that our control rules reduce waiting times by 1025% compared to
a straightforward periodic rule. Dynamic programming yields the best results in terms of mean waiting
times.
2003 Elsevier Ltd. All rights reserved.
Keywords: Trac control; Simulation; Transportation networks; Heuristics

1. Introduction
In various situations, trac from two directions shares the same infrastructure. Well-known
examples are roadblocks caused by trac accidents or road maintenance (Son, 1999). We encountered a similar setting in a project on the design of an underground transportation system
near Schiphol Airport in the Netherlands (van der Heijden et al., 2002a). Here Automatic Guided

Corresponding author. Tel.: +31-53-4893893; fax: +31-53-4892159.

E-mail address: m.j.r.ebben@utwente.nl (M. Ebben).

0191-2615/$ - see front matter 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0191-2615(03)00075-4

442

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

Fig. 1. A two-way tube connecting double tube systems.

Vehicles (AGVs) carry freight through underground tubes between various terminals. To reduce
infrastructure investment, some terminals are connected by a single tube for trac in both
directions. Intelligent control of the driving direction is required to guarantee acceptable order
throughput times. Although this problem is related to traditional trac systems, an essential
dierence is that the behaviour of AGVs can be directly inuenced. Moreover, quite precise information is available on expected arrival times of AGVs at the two-way track entrance, because
the driving behaviour of AGVs is more predictable than that of passenger cars and because routes
are known in advance. To exploit this information, we develop new control rules, focusing on
delay reduction.
We address a tube allowing for alternating trac as a two-way tube, see Fig. 1 for an example
with connections to double tube systems at both sides. As is shown in the gure, all AGVs have an
identical length, but some of them may be driving in a convoy with minimal mutual distance.
According to a control rule, AGVs arriving at the two-way tube may either enter immediately or
queue until a signal is given that access is allowed. Each control rule generates the system states,
shown in Table 1, in cyclic order. We use the term green time for the time spent in the states b and
d and the term clearance time for the time spent in the states a and c. The clearance time refers
to the period with a red trac light at both ends, while AGVs still have to leave the two-way tube.
The cumulative time spent in the four system states makes up the cycle time. The time spent in
each system state depends on the driving time in the two-way tube, the arrival distributions at
both sides and the control rule. The starts of the red light intervals a and c can be considered as
decision variables in the operational control. We assume that at each moment t it is known how
many vehicles are present in the two-way tube and that an end-of-clearance signal is available.
Technologically, this can be accomplished by using induction wiring at both ends of the tube.
Similarly, information on future arrivals can be obtained from induction wiring at a certain
distance from both tube entrances. For example, with passage loops at a reasonable distance from
the entrance and given the constant speed of the vehicles, we can derive the number of vehicles
that will be waiting at the entrance.
The purpose of a control rule is to minimise average vehicle delay. Note that we consider the
system as a discrete process, as opposed to what is often assumed in trac literature. A
straightforward control rule is First-Come First-Served (FCFS). However, this rule will only lead
Table 1
Description of the system states
State

Description

AGVs from the right have to wait, but AGVs from the left cannot yet enter the tube, because it still
contains AGVs driving from right to left
AGVs from the left can enter and AGVs from the right have to wait
AGVs from the left are stopped and AGVs from the right wait until the tube is cleared
AGVs from the right can enter and AGVs from the left have to wait

b
c
d

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

443

to a reasonable performance in the case of very low trac intensity and relatively short driving
times along a two-way section (for example crossroads). This is not true for the Schiphol system,
where a high trac intensity and long tube driving times are envisaged (510 min). A second
obvious option is switching the driving direction periodically; the system spends a xed time in
both the system states a&b and c&d. Note that the eective green time is not xed, because it
depends on the preceding clearing time. AGVs cannot enter the tube if AGVs from the other
directions are still in the tube. The advantage of a periodic control rule lies in its simplicity; it
requires no information on system status. An obvious drawback is that information on AGV
arrivals and queues at both two-way tube entrances is not used. In particular, when the trac
intensities at both sides of the two-way tube uctuate in time (as is true in our application), a
simple periodic control rule may lead to excessive waiting times.
A third simple rule is to serve the entire queue in one direction rst, and then the entire queue in
the other direction. In a static situation without any information on future arrivals, it can be
argued that this convoy rule leads to an optimal solution. However, a simple example shows that
this approach need not to be optimal in a dynamic context when new AGVs arrive while both
queues are processed. Imagine the following situation at some time t. The last vehicle from the
right has entered the tube 0.6 min ago and 10 AGVs are waiting at the left side. The driving time
through the tube equals 6 min, so the AGVs from the left could start driving at t 5:4; the AGVs
enter the tube at a rate of 20 vehicles per minute. Another AGV arrives at the right side of the
tube; no other AGVs arrive. If the latter AGV has to wait, the additional waiting time incurred
equals 5.4 (remaining clearance time when the AGV arrives) + 10/20 (convoy at the left enters the
tube) + 6 (passing time of the convoy) 11.9 min. On the other hand, if the AGV from the right is
allowed to enter the tube, the additional waiting time incurred equals 0.6 min for each AGV
waiting at the left side, so 10
0:6 6 min. This example shows that the convoy rule is not
optimal for this case.
Therefore, we developed adaptive control rules with increasing complexity and information
usage (Sections 3 and 4). We focus on control rules that are suitable for on-line usage: they have to
be ecient in terms of computation time. The potential of the new control rules is demonstrated
by an extensive simulation study in a rolling horizon setting (Sections 5 and 6). We will also
discuss the implications of embedding the two-way tube in a closed transportation network
(Section 7). In all cases, we use periodic control and the convoy rule as benchmarks.

2. Literature review
Two-way track control is related to junction control supported by trac signals, see Newell
(1988) for an overview. In the early days trac signals were usually scheduled according to a predetermined scheme (cf. Bell, 1992). A periodic control rule is a good example of such a scheme.
Newell (1988) gave a relation between the mean waiting times in a deterministic and a stochastic
setting, assuming that the clearance time is deterministic. He showed that a deterministic
approach suces if the clearance time goes to innity, but also that this is not applicable for most
realistic settings. Despite the long clearance times in our application, the dierence in delay between the stochastic and deterministic case is still around 50% (Newell, 1988, Eq. 2.3.12), apart
from the fact that we do not face deterministic clearance times. Note that in the analysis of

444

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

intersections the vehicles mostly are regarded as a continuous uid (cf. Newell, 2002). In our
dynamic control rules we exploit the fact that in reality it is a discrete arrival process. In case
arrival rates are not too high, improvements can be expected from dynamic control rules.
Stochastic models for periodic control have been addressed by several authors. Mung et al.
(1996) derived distributions of queue lengths at xed time trac signals. Heidemann (1994) derived analytical results on statistical distributions of queue lengths and delays at trac signals,
given Poisson arrivals. He compared these results with several approximations (Webster, 1958;
Miller, 1968). Hu et al. (1997) extended Heidemanns model to the multi-lane case, where multiple
vehicles may enter the trac intersection simultaneously. As is argued in van der Heijden et al.
(2002b), these models are not applicable to the two-way tube problem, because the eective green
period and the eective red period are not constant and known. Therefore, van der Heijden et al.
(2002b) developed an approximation for the mean waiting time in periodic systems with random
clearance times. Still, none of the papers above exploit the information on the specic system state
(AGVs in queue and en route to the two-way tube) to reduce average waiting times. This is not
surprising, because the availability of such information is usually limited. As explained in the
introduction, the AGV systems that we have in mind can provide this additional information.
Situations analogous to our two-way tube arise in trac control, when one lane of a two-lane
highway is closed because of maintenance or construction activities. Ceder and Regueros (1990)
and Ceder (2000) used a periodic model and delay formula based on Webster (1958) as input for
an optimisation model determining the lane closure policy and the lane closure length. Cassidy et
al. (1994) expanded a deterministic equation for the delay using a statistical analysis of empirical
data. An explicit control policy has not been dened. Shibuya et al. (1996) optimised green intervals using a regression equation for the delay. They observed discrepancies with simulated
green intervals and concluded that eective control is very dicult under heavy trac ows. In
Son (1999), expressions for the mean delay have been derived, based on the assumption that
vehicles arriving within a period H after a queue has entered the two-way track are allowed to
enter the two-way track as well (green time extension). Here the value of H has been estimated
empirically using discrete choice techniques. All these papers do not focus on waiting time
reduction using an adaptive, state dependent control policy.
Finally, we note that problems rather similar to the two-way tube problem have been studied in
the machine scheduling literature, particularly in choosing batch sizes for ovens, see e.g. Uzsoy et
al. (1992, 1994), Webster and Baker (1995), van der Zee et al. (1997) and Potts and Kovalyov
(2000). Similarities with the two-way tube system can be found in the fact that goods (AGVs) have
to be batched, with the processing time of a batch equal to the driving time on the two-way
track. An essential dierence is the fact that processing times are xed for oven systems, because
they are related to static product and process characteristics. However, the processing time of a
convoy in a two-way tube depends on the time between the rst and the last AGV in a convoy.
Furthermore, oven systems set restrictions on the number of products in a batch, but there is no a
priori limit on the number of AGVs in a convoy.
Summarising, the most important dierences compared to literature are:
(a) Signicant throughput time in the two-way tube (several minutes).
(b) Focus on adaptive, state-dependent control policies, using information on queue length and/
or future AGV arrivals.

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

445

(c) Embedding adaptive control rules in a rolling horizon environment.

(d) No maximum convoy (batch) size.

3. Assumptions and control rule concepts

3.1. Assumptions
Our models are based on the following assumptions:
(1) All AGVs are identical.
(2) When driving, the AGVs travel at the same, constant speed.
(3) Queuing AGVs accelerate instantaneously to their normal speed, when activated, and AGVs
can stop instantaneously when arriving at a queue.
(4) AGVs waiting in queue enter the shared lane with a xed minimum succession time (d).
Instantaneous acceleration and deceleration (assumption 3) is obviously not realistic, but the
impact on the system performance is low if little time is required compared to the driving time in
the two-way tube. Furthermore, it is possible to extend our methods to include the acceleration/
deceleration time. In practice, the succession time d (assumption 4) will be strictly positive because
of the length of an AGV; two vehicles cannot enter a single lane simultaneously. Also, safety
margins can be included to minimise the collision probability, taking into account the deceleration. Note that we do not make an explicit assumption about the arrival processes, both regarding
inter-arrival distribution and lot-size (single or batch arrivals).
In the remainder of this section we characterise the rules for real-time two-way tube control in
terms of information usage and decision structure. Mathematical details are described in Section 4.
3.2. Characterisation of control rules
For the construction of decision rules we chose minimum average waiting time at the two-way
tube as criterion. The waiting time is dened as the time from the arrival of an AGV at the entrance of the two-way tube until it enters the two-way tube. Still, the basic ideas presented in this
paper can also be applied to some other criteria, such as minimum lateness. Our decision rules
may use information on the system state at some decision moment t0 . We distinguish three options
for information availability:
(a) No information on AGV queues and arrivals, only average AGV arrival intensities are
known.
(b) Only local information on queue lengths at the two-way tube entrances is available.
(c) Prior information, i.e. both on current queue lengths and on future arrivals within a certain
information horizon H > 0, is available.
At each decision moment the allowed decisions are: let the green trac light remain green until
the next decision moment or put the green light on red so that a change of direction is enforced

446

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

Table 2
Classication of the control rules
Decision option
One change
Multiple changes

Information availability
No information

Local information

Prior information

Periodic Control (PCR)

Convoy rule (Conv)

Adaptive local control

Adaptive look-ahead control
(ARloc)
(ARla)
Dynamic Programming (DP)

immediately after this decision moment. Policies open to the controller are either to take a single
decision on the timing of the next direction change, or to make a schedule of multiple direction
changes within the information horizon H , i.e. all known AGV arrivals are included in the
schedule.
In Table 2 we classify our control rules according to the information availability and the decision options. If no information on the AGVs waiting or arriving is available, a periodic control
rule and a convoy rule are realistic strategies. The convoy rule does not need any parameter setting;
the direction is changed directly after the last vehicle of a convoy has entered the two-way tube.
The xed time interval P of the periodic control rule can be based on o-line optimisation,
considering the historical and/or expected trac ows (cf. van der Heijden et al., 2002b). In the
other cases a decision policy has to be constructed.
A simple adaptive local control rule (Section 4.1) uses information on queue lengths only to
make a decision on the next direction change. Multiple changeovers can be determined by a
dynamic programming approach that yields an optimal solution in a static situation. If only local
information on queue lengths is available, obviously the whole queue at the right (left) rst passes
the two-way tube if it is decided that the direction is changed to rightleft (leftright). Any other
option, for example processing the queue at the right in two separate batches, is inferior in terms
of total waiting time for the vehicles in queue because of the additional clearance time in case of
an additional direction change. For this reason, the dynamic programming approach naturally
leads to batch processing, equivalent to adaptive local control; the entire queue present at the
decision moment enters the two-way tube.
We consider two control rules based on information on future arrivals. First, we develop an
adaptive look-ahead control rule for determining the next direction change (Section 4.2). Second, a
dynamic programming rule is introduced (Section 4.3), which computes an optimal sequence of
direction changes. Because it is well known that an optimal algorithm for a static situation may be
inferior to a simple heuristic in a rolling horizon setting, we analyse both options in the remainder
of this paper. In a dynamic situation, information changes when additional AGVs arrive and/or
when forecasts of future AGV arrivals are improved. As time proceeds, new information becomes
available and rescheduling is necessary. This implies that the rescheduling frequency in a rolling
horizon approach appears as an additional parameter. Regarding decision policies that determine
multiple changeovers, it is natural to only implement the rst changeover in the sequence, taking
future direction changes into account. The other decisions will be overruled by information updates.
Finally, we note that we restrict the decision moments to the following list of discrete
events:

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

447

(I) the end of state a, where the two-way tube is cleared for trac from the left;
(II) the arrival of a new AGV at the left or the right end of the two-way tube in state b, with a
green trac light at the left entrance;
(III) the end of state c, where the two-way tube is cleared for trac from the right;
(IV) the arrival of a new AGV at the left or the right end of the two-way tube in state d, with a
green trac light at the right entrance.
Hence, there are no decision moments during clearance periods (states a and c); we do not
reconsider a decision to change the driving direction.

4. Control rules
Here we address the mathematical details of the adaptive control rules as introduced in Section 3.2. For simplicity, we rst discuss each control rule assuming that the minimum succession
time d 0. Next, we generalise our results to d > 0. We use the following basic notation:
decision moment
information horizon, i.e. at a decision moment t0 , AGV arrivals up to t0 H are known
the clearance time of the two-way tube at t0 , i.e. the time needed for the last AGV to leave
the two-way tube from time t0 on if no additional AGVs enter the two-way tube
the ith arrival at the right (left) side of the two-way tube, where tir til 6 t0 if an AGV has
tir ; til
already arrived
qr t; ql t the number of AGVs in queue at the right (left) at time t

t0
H
Ct0

Note that the clearance time Ct0 depends on the AGV positions in the tube at time t0 and is in
general not equal to the driving time in the tube T .
4.1. Adaptive local control (ARloc)
Only the queue size at either side of the two-way tube is known. Without loss of generality, we
assume that the current driving direction is from left to right. Recall that we only need to consider
two options (Section 3.2): either the queue at the left passes the two-way tube rst or the queue at
the right passes the two-way tube rst. Note that in a dynamic situation this is substantially
dierent from the convoy rule, because of additional decision moments caused by new arrivals
after convoy entrance. Our optimisation criterion is minimum average waiting time. In a dynamic
situation, this means that we should only account for additional waiting times that can still be
inuenced by the decisions to be taken. Waiting times already incurred at the decision moment are
not relevant anymore. Therefore, our adaptive control rules minimise a cost function equal to the
total additional waiting times incurred by all AGVs to be scheduled. Still, we will measure the
average delay at the two-way tube when evaluating the control rules in our simulation study
(Section 6).

448

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

4.1.1. The case d 0

A direction change means that the AGVs at the right side will pass the two-way tube rst. This
will induce waiting time for the AGVs in queue at the left, being at least equal to the clearance
time (Ct0 ) plus the driving time (T ). So the costs caused by a direction change are
Wchange ql t0 Ct0 T

The alternative is that the direction is not changed immediately, but after the AGVs from the left
have entered the two-way tube. Then, all AGVs in queue at the right (qr t0 ) incur additional
waiting time, equal to the dierence between the clearance time before making a decision (Ct0 )
and the new clearance time caused by AGVs entering the two-way tube from the left ( the
driving time T ):
Wnochange qr t0 T  Ct0

The direction is changed immediately when Wchange < Wnochange . Otherwise, in a static situation the
direction change is planned after all AGVs in the queue at the left entrance have entered the twoway tube. In a dynamic situation, it will be natural to make a new decision when the next AGV
arrives, which makes it dierent from the convoy rule.
4.1.2. The case d > 0
Eqs. (1) and (2) have to be extended to account for the time it takes a convoy to enter the twoway tube. Except for the rst AGV, all AGVs have to remain at a distance d from their predecessor:
Wchange ql t0 Ct0 T d  maxfqr t0  1; 0g

Wnochange qr t0 T  Ct0 d  maxfql t0  1; 0g

4.2. Adaptive look-ahead control (ARla)

Look-ahead control is an extension of local control, taking into account future arrivals within
the information horizon H. Whereas local control leads to a simple choice between two options
(either the convoy from the right or from the left has priority), we can now change the driving
direction at the decision moment or after the ith arrival from the left, see the example in the
introduction.
4.2.1. The case d 0
The costs related to a direct direction change are a straightforward extension of formula (1):
X
Wchange
minft0 T Ct0  tkl ; T Ct0 g
5
tkl 6 t0 T Ct0

Note that this equation reduces to (1) when tkl t0 . The formula includes the waiting time of
arrivals (tkl ) at the left entrance of the two-way tube up to t0 Ct0 T unless, given the
restrictions set on the information horizon H , this information is not available. Note how (5) also
covers those situations in which decision making does not coincide with the arrival of an AGV
(tkl < t0 ).

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

449

The alternative is to schedule the direction change after the ith AGV entering the two-way tube.
At that time qr til T AGVs are queuing at the right entrance. For each option costs are computed as
X
X
til 2T  tkl
til T  maxftkr ; t0 Ct0 g
6
Wnochange;i
til <tkl <til 2T

tkr <til T

The rst term considers the waiting times of AGVs arriving at the left side of the two-way tube
within the interval til ; til 2T , i.e., after i AGVs have entered the two-way tube and before the
scheduled AGVs from the right have passed the two-way tube. The second term of (6) represents
waiting times for AGVs in the queue and arriving at the right side of the two-way tube up to a
time horizon til T .
A straightforward comparison between the options is unfortunately not possible, because we
compute waiting times for dierent schedules. In formula (5), we take into account all waiting
times incurred until t0 T Ct0 , whereas in (6) we use all waiting times up to time til 2T .
These numbers represent the planning horizon in the various decision options. Obviously, the total
additional waiting time increases with the length of the planning horizon. For fair comparisons,
we weigh the waiting times with the length of the planning horizon, thereby obtaining the additional waiting time incurred per time unit. Weighted costs associated with each option are
formulated as



Wchange
Wnochange;i
and WCnochange min l
7
WCchange
i
T Ct0
ti 2T  t0
The direction is changed immediately if WCchange < WCnochange , otherwise the changeover is planned
after the ith arrival which causes the minimum WCnochange . Note that this planning can be overruled by later decisions.
4.2.2. The case d > 0
Again, the additional waiting time caused by the required distance between AGVs has to be
added:
X
minft0 T Ct0 d  qr t0 Ct0  1
Wchange
tkl 6 t0 T Ct0 dqr t0 Ct0 1

 tkl ; T Ct0 d  qr t0 Ct0  1g

X
Wnochange;i

8
maxftil ; t0 d  i  1g

til <tkl <maxftil ;t0 di1g2T dqr maxftil ;t0 di1gT 1

2T  tkl d  qr maxftil ; t0 d  i  1g T  1
X

maxftil ; t0 d  i  1g T  maxftkr ; t0 Ct0 g

tkr <maxftil ;t0 di1gT

We assume that the number in the queue is greater than 0. Otherwise we have to take the maximum of qr   1 and 0, as is shown in (3) and (4). The number of options may increase as a result
of the entrance time. The weighted costs associated with each option are formulated as

450

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

Wchange
T Ct0 d  qr t0 Ct0  1



Wnochange;i
min
i
maxftil ; t0 d  i  1g 2T  t0 d  qr maxftil ; t0 d  i  1g T  1

WCchange
WCnochange

10
4.3. Dynamic programming (DP)
For those situations where multiple direction changes have to be scheduled, we propose a
Dynamic Programming approach. The scheduling decision involves nding the sequence of
convoys which minimises total additional waiting time. Denoting by NL (NR ) the number of AGVs
from the left (right) within the information horizon at t0 , we can formulate the objective function
as
MINffn NL ; NR ; t g

11

nL;R

Here fn i; j; t is dened as the minimum total additional waiting time for all AGVs present at the
two-way tube, or arriving within the horizon H, if at time t already i AGVs from the left and j
AGVs from the right have passed the two-way tube (i 0; 1; . . . ; NL ; j 0; 1; . . . ; NR ), and the
last convoy passing the two-way tube came from direction n (n L; R). Hence t is a point in time
at which all AGVs, known at t0 , have been processed. Note that t is just an auxiliary variable for
the recursion.
4.3.1. The case d 0
Given the initial conditions at the decision moment (t0 ): fn 0; 0; t0 0, the cost function fn 
can be formulated, starting with the left side, as
fL i; j  k2 ; t0 min fci  k1 ; j  k2 ; t; i; j  k2 ; t0  fR i  k1 ; j  k2 ; tg
k1 1;...;i

00

fR i; j; t min fci; j  k2 ; t0 ; i; j; t00  fL i; j  k2 ; t0 g

k2 1;...;j

with ci  k1 ; j  k2 ; t; i; j  k2 ; t 

yi
X

maxft  tyl ; 0g

yik1 1

and ci; j  k2 ; t0 ; i; j; t00 

yj
X

12

maxft0  tyr ; 0g

yjk2 1
0

where t T

maxft; til g

and t00 T maxft0 ; tjr g

We assume that a convoy contains at least 1 AGV, without aecting optimality, because otherwise
there is no reason to change the direction. The function ci  k1 ; j  k2 ; t; i; j  k2 ; t0  computes
waiting times for a convoy of k1 AGVs that has entered the two-way tube at time t and left it at
time t0 . This time is determined by the question of whether the last AGV arrived during the red
(til 6 t) or green period (til > t). The computation times do not necessarily increase enormously with
more arrivals, because not all discrete points in time have to be computed; the algorithm takes

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

451

long time steps of length T . AGVs that wait in queue when the trac light becomes green will all
enter the two-way tube. It is not optimal to cut a queue into two separate convoys.
4.3.2. The case d > 0
In this case only the cost functions and the decision times change, the recursion remains unchanged.
ci  k1 ; j  k2 ; t; i; j  k2 ; t0 

yi
X

maxft  tyl d  y  i k1  2; 0g

yik1 1
yj
X

ci; j  k2 ; t0 ; i; j; t00 

maxft0  tyr d  y  j k2  2; 0g

13

yjk2 1
0

where t T maxft d
ql t  1; til g and t00 T maxft0 d  qr t0  1; tir g

In the implementation of the algorithm we assume that all AGVs that are waiting in the queue
when the trac light switches to green enter the two-way tube in one convoy. A queue is not
separated into two convoys, which might happen for a d > 0. This assumption signicantly reduces computation times.

5. Design of the simulation study

To demonstrate the potential of the strategies for two-way tube control, we set up an extensive
simulation study. We study the behaviour of a two-way tube in an open system (Section 6) as well
as in a closed system (Section 7). The dynamic behaviour in a closed system is hard to predict,
because convoys created by the two-way tube may cause batch arrivals of the return ow of AGVs
at a later point in time. We use the Schiphol-case to conduct simulation experiments for closed
systems.
The design of the simulation study is based on gures from the Schiphol-case. In Table 3, the
experimental factors and their ranges are shown. Note that these settings with respect to arrival
rates and tube length are not substantially dierent from Cassidy et al. (1994). For convenience,
the control rules are abbreviated as Conv (Convoy rule), PCR (Periodic Control Rule), ARloc
(Adaptive Rule with local control), ARla (Adaptive Rule with look-ahead), and DP (Dynamic Programming). The experimental factors concern the arrival patterns of AGVs (interval,
Table 3
Experimental factors
Factor

Range

A. Control rule
B. Average inter-arrival time (min)
C. Arrival distribution
D. Lot size
E. Minimum succession time d (s)
F. Two-way tube length (m)

Conv, PCR, ARloc, ARla, DP

0.25, 0.33, 0.5, 1, 1.5, 2, 4
Poisson, Uniform
1, Uniform(3, 9)
0, 3.5
1000, 2000, 3000

452

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

distribution and lot size) and two-way tube length. The driving speed of the AGVs is constant and
equals 6 m/s. For each control rule average waiting time per AGV has been measured for default
settings. These results are shown in Section 6.1. Next, alternative system congurations were
chosen by changing the value of only one of the factors CF (Section 6.2). Also, we examine the
sensitivity of our results to two important control parameters, the information horizon and the
planning frequency (Section 6.3).
Note that we determine the xed time interval P of the periodic control rule by using the
formulas derived in van der Heijden et al. (2002b). This results in change-over times of 5.9, 6.1,
6.3, 6.8, 7.1, 7.3 and 8 min corresponding to average inter-arrival times of 0.25, 0.33, 0.5, 1, 1.5, 2
and 4 min, respectively. For the adaptive look-ahead rule (ARLA ), we chose to consider all options
for changing the driving direction within the interval t0 ; t0 T . Many choices are possible here;
our choice of the interval was motivated by two arguments. First, considering system characteristics, the driving time in the two-way tube forms a natural horizon. Secondly, as a result of
simulation experiments, we found that the choice of the interval is an acceptable trade-o between
the information requirement and the performance improvement. For the information horizon, we
take as initial setting H 3T . This period logically follows from the above choice to consider all
possible moments for changing the two-way tube direction up to once the driving time in the twoway tube (T ). Given a maximum clearance time of T , waiting times for AGVs may be inuenced
up to 3T . Recall that we will examine the sensitivity of our results to the value of H in Section 6.3.
In our experiments we adopted the batch means method (cf. Law and Kelton, 2000). Each
batch equals one day. The rst batch is discarded to account for any start-up bias. We chose the
number of batches such, that the relative simulation error is 1% at most (within a 95% condence
interval).

6. Analysis of simulation results

6.1. Default settings
The default settings concern single AGVs arriving according to a Poisson process at a 2000meter long two-way tube. Obviously, the average waiting time strongly depends on the AGV
arrival interval and tube length. In order to facilitate the comparison of the results, the performance for each setting is indicated in terms of normalised average delay, i.e., the average waiting
time at the two-way tube divided by the driving time in the two-way tube (T ). Since we divide time
by time normalised average delay has no dimension.
Fig. 2 shows that the average waiting time of the convoy rule is equal to the driving time in the
tube, as expected from theory, but only in case of very high arrival rates are the results acceptable.
With very high arrival rates it seems that it is no longer benecial to wait for additional arrivals,
because this causes too much waiting time to waiting vehicles at the other end of the tube. In this
case a convoy rule might be preferred because of its simplicity. The adaptive rules clearly beat the
PCR-rule by percentages up to 25% at low arrival rates. The dierences between the adaptive rules
are smaller. The look-ahead rule performs 23% better than the local control rule, while the
dynamic programming rule performs 58% better than the look-ahead rule. The outcomes clearly
conrm the general proposition that the more information on future arrivals is included in de-

normalized average delay

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

1.1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3

453

PCR
Conv
ARloc
ARla
DP

0.5

1.5

2.5

3.5

arrival interval (minutes)

Fig. 2. Normalised average delay for the default settings (simulation results).

cision-making, the better the results. Note that the performance dierences presented in Fig. 2
may all be considered signicant, given a paired t-test with signicance level 95%.
A drawback of a dynamic programming approach might be the computational eort. A series
of experiments indicated that even in the worst case evaluated in the simulation study (see Section
6; average inter-arrival time of 0.25 min), the dynamic programming rule requires less than a
second per changeover using a Pentium III 500 MHz. Although computation times for the dynamic programming rules tend to increase exponentially for higher arrival rates, this outcome still
leaves a lot of room for practical application.
6.2. Alternative system congurations
As alternative system congurations, we simulated two-way tubes with the other parameters
values from Table 3. Here we briey discuss our main ndings. Regarding two-way tube length,
we found that this factor hardly inuences the relative performance of the control rules. The
inuence of the arrival distribution on system performance is studied in two ways. In the rst
series of experiments the Poisson process with a negative exponential distribution is replaced by a
more regular Uniform distribution. Secondly, we considered compound Poisson arrivals, where
the lot-size of arriving AGVs is uniformly drawn from the interval 3; 9. Note that we increased
the average inter-arrival times (3, 6, 9 and 12 min) correspondingly.
Two conclusions can be drawn from this series of experiments:
1. The greater the irregularity of the arrival pattern, the better the adaptive control rules perform
in comparison to the periodic control rule. This is due to their responsiveness to system state.
2. More irregular arrival patterns lead to lower normalised average delay. More frequently, large
time gaps occur between AGVs, which are eciently used by the adaptive rules for changing
the two-way tube direction.
These conclusions are illustrated in Fig. 3. The dierence between the performances of the periodic control rule and the dynamic programming rule is much smaller for Uniform arrivals, than it
is for Poisson arrivals, usually less than 6%. We also see that normalised average delay is much
smaller for Poisson arrivals, than it is for Uniform arrivals (cf. DP Poisson and DP Uniform).

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

normalized average delay

454

1
0.9
0.8

PCR Uniform

0.7

DP Uniform

0.6

PCR Poisson
DP Poisson

0.5
0.4
0.3
0

0.5

1.5

2.5

3.5

arrival interval (minutes)

normalized average delay

Fig. 3. Periodic control and dynamic programming for Uniform and Poisson arrivals (simulation results).

1.6
1.4
1.2

PCR

Conv
ARloc

0.8

ARla

0.6

DP

0.4
0.2
0

0.5

1.5

2.5

3.5

arrival interval (minutes)

Fig. 4. Intermediate distances and AGV length considered, d 3:5 s (simulation results).

Similar conclusions were found by Glassey and Weng (1991) in the context of dynamic machine
scheduling.
In the previous experiments we considered the symmetrical case. However, if the arrival rates
from the left and right side dier (asymmetric case), we expect that the performance of the periodic control rule will be worse. This hypothesis was conrmed in an additional experiment.
Given an average inter-arrival time of 0.5 min from one side and 1 min for the other side, the
relative dierence between the periodic control rule and the dynamic programming rule rises to
22% (cf. Fig. 2).
Finally, we examine the impact of the succession time d. Fig. 4 presents results for d 3:5 s,
which is the actual value in the Schiphol-case. For the periodic rule, we use switching periods of
P 11 12 ; 9 12 ; 8 14 ; 7 12 ; 7; 6 12 ; and 6 14 min corresponding to k equal to 0.25, 0.33, 0.5, 1, 1.5, 2
and 4 min (cf. van der Heijden et al., 2002b). The remaining settings correspond to the default
settings.
The results in Fig. 4 are similar to those in Fig. 2, although we clearly see the eect of safety
precautions in terms of longer waiting times in the case of high trac intensities. Again, adaptive
rules perform signicantly better than both the periodic rule and the convoy rule. It is remarkable
that most of the performance improvement is already obtained by local control using information
on queue sizes only.

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

1

0.9
0.8
DPconvoy(1)
0.7

DP(1)

0.6

ARla(1)

0.5

normalized average delay

normalized average delay

455

0.9
0.8

DPconvoy(4)

0.7

DP(4)

0.6

ARla(4)

0.5

0.4
1000 2000 3000 4000 5000 6000 7000 8000

0.4
1000 2000 3000 4000 5000 6000 7000 8000

information horizon (meter)

information horizon (meter)

Fig. 5. Sensitivity for the length of the information horizon and planning frequency for k 1 (left gure) and k 4 min
respectively (simulation results).

6.3. Sensitivity for the information horizon and planning frequency

An interesting question from a practical perspective is how the performance of the new control
rules relates to the availability of information on future arrivals. To answer this question, we
study the relationship between the length of the information horizon and system performance. In
this section, we express the information horizon in meters rather than in time, because this naturally relates to facilities for vehicle detection like sensors or inductive loops. The experiment
includes the DP-rule and the ARLA rule. As far as the DP rule is concerned, two scheduling
frequencies are considered: per arriving AGV (DP as considered in Section 4.3) and per convoy
(DPconvoy). The scheduling frequency for the DPconvoy rule is related to a planned direction
change, i.e. no new decision is made before the planned convoy has left the two-way tube. The
two-way tube length is set to 2000 meters, the succession time is set to d 0 and AGVs arrive
according to a Poisson process with average arrival intervals of 1 and 4 min.
Fig. 5 shows the numerical results, where the average inter-arrival time in minutes is given
between parentheses. We see that the dynamic programming rule (DP) and the look-ahead rule
(ARLA ) are not very sensitive to the length of the information horizon. An information horizon of
3000 meters already gives good results, while more distant information horizons do not lead to
considerable improvement. The fact that performance is most inuenced at low arrival rates can
easily be explained by the relative weight of information on arriving AGVs under these circumstances. Another eect, which is clearly shown by Fig. 5, is the eect of the planning frequency in a
dynamic context; compare the results for both DP rules. Clearly, a lower planning frequency can
signicantly reduce system performance. For longer information horizons, the eect is smaller.
7. A two-way tube in a closed system
7.1. The eects of convoys on the system
Until now we have studied the two-way tube in isolation, as part of an open system. As a
consequence, arrivals were considered independently of each other and an innite population was
assumed. It is reasonable to assume an open system in the case of a road network or a large
transportation network. In these cases, the convoys that are formed by the two-way tube are

456

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

dispersed to dierent destinations and the population is also large compared to the trac in the
section under consideration. In a closed system these assumptions can no longer be justied,
making it necessary to study the eects of interaction with other elements in the network (compare
the attention paid in the trac literature to co-ordinating trac signals, see e.g. Bell, 1992). For
example, let us consider the situation where the network at one side of the two-way tube is a single
terminal. The convoys that leave the two-way tube drive to the terminal, load/unload, and
probably soon return to the same two-way tube. So the two-way tube will face batch arrivals of
AGVs, where the batch size is heavily inuenced by a previous direction change. Decisions on
two-way tube control may therefore have serious impact on future decisions, as well as having a
signicant impact on total system performance.
7.2. Numerical results for the Schiphol-case
To check whether our conclusions for two-way tube control in an open system are equally
valid for a closed system, we embedded the rules in a simulation model for the Schiphol system.
The layout that was studied contains one two-way tube (see Fig. 6). At one side of the two-way
tube there is only one terminal. At the other side the two-way tube is connected to several
terminals. The two-way tube has a length of 1500 m and for both sides accurate information on
arrivals is available by recording AGV movements 1500 m before the entrance. The required
distance between two AGVs is 3.5 s. The control rules still aim to minimise the waiting
time locally at the start of the two-way tube, but we measure the total system performance in
terms of the service level, i.e. the percentage of transportation jobs that is handled before the due
time.
The question is whether the local control rules that perform best in minimising the average
waiting time at an isolated two-way tube also realise the best system performance. For PCR
after some initial experiments we chose a switching period of 7 min. The simulation results
conrm the ranking of the control rules with respect to our performance measure, which can be
seen in Table 4.

Fig. 6. Layout of the Schiphol system.

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

457

Table 4
Overall service level (% on-time delivery) of the Schiphol system using dierent control rules
Control rule

180 AGVs

190 AGVs

PCR (7 min)
ARLOC
ARLA
DP

90.2%
92.4%
93.5%
99.1%

99.7%
99.5%
99.6%
100%

8. Conclusions
In this paper we addressed the issue of real-time control of a two-way tube. Several new
control rules were dened as an alternative to the classic periodic control rule. The adaptive
control rules try to use local information on queue lengths and look-ahead information on
future arrivals to improve on system performance. Simulation results indicated that, depending
on the amount of information available, adaptive control rules improve system performance up
to 10% for high arrival rates and up to 25% for low arrival rates. Best performance was realised
by a rule based on dynamic programming. While it is known that a dynamic programming
approach may involve high computational costs, simulation results for a realistic large-scale
transportation network indicated that computation times for the dynamic programming rules
are acceptable for real-time applications. In general, the performance of the rules was not much
inuenced by the length of the information horizon. An experiment that included the twoway tube in a closed network, the Schiphol system, showed that our conclusions still hold
for a closed system. An interesting topic for future research is a more thorough examination
of the interaction eects with terminals or other two-way tubes within an open or closed network.

References
Bell, M.G.H., 1992. Future directions in trac signal control. Transportation Research A 26, 303313.
Cassidy, M.J., Son, Y.T., Rosowsky, D.V., 1994. Estimating motorist delay at two-lane highway work zones.
Transportation Research A 28, 433444.
Ceder, A., Regueros, A., 1990. Trac control (at alternate one-way sections) during lane closure periods of a two-way
highway. In: Koshi, M. (Ed.), Proceedings of the Eleventh International Symposium on Transportation and Trac
Theory, pp. 551568.
Ceder, A., 2000. An application of an optimal trac control during lane closure periods of a two-lane road. Journal of
Advanced Transportation 34, 173190.
Glassey, C.R., Weng, W.W., 1991. Dynamic batching heuristics for simultaneous processing. IEEE Transactions on
semiconductor manufacturing 4, 7782.
Heidemann, D., 1994. Queue length and delay distributions at trac signals. Transportation Research B 28, 377389.
Law, A.M., Kelton, W.D., 2000. Simulation Modelling and Analysis, third ed. McGraw-Hill, New York.
Miller, A.J., 1968. Settings for xed-cycle trac signals. Operations Research Quarterly 14, 373386.
Mung, G.K.S., Poon, C.K.A., Lam, W.H.K., 1996. Distributions of queue lengths at xed time trac signals.
Transportation Research B 30, 421439.
Newell, G.F., 1988. Theory of highway trac signals. Report no. UCB-ITS-CN-89-1, Institute for Transportation
Studies, University of California.

458

M. Ebben et al. / Transportation Research Part B 38 (2004) 441458

Newell, G.F., 2002. The oversaturated isolated intersection. In: Transportation and Trac Theory in the 21st Century:
Proceedings of the 15th International Symposium on Transportation and Trac Theory, Adelaide, Australia, pp.
119135.
Potts, C.N., Kovalyov, M.Y., 2000. Scheduling with batching: a review. European Journal of Operational Research
120, 228249.
Shibuya, S., Nakatsui, T., Fujiwara, T., Matsuyama, E., 1996. Trac control at agger-operated work zones on twolane roads. Transportation Research Record 1529, 39.
Son, Y.T., 1999. Queueing delay models for two-lane highway work zones. Transportation Research B 33, 459471.
Uzsoy, R., Lee, C.Y., Martin-Vega, L.A., 1992. A review of production planning and scheduling models in the
semiconductor industry, Part I: System characteristics, performance evaluation and production planning. IIE
Transactions 24, 4760.
Uzsoy, R., Lee, C.Y., Martin-Vega, L.A., 1994. A review of production planning and scheduling models in the
semiconductor industry, Part II: Shop-oor control. IIE Transactions 26, 4455.
van der Heijden, M.C., van Harten, A., Ebben, M.J.R., Saanen, Y., Valentin, E., Verbraeck, A., 2002a. Using
simulation to design an automated underground system for transporting freight around Schiphol Airport. Interfaces
32, 119.
van der Heijden, M.C., van Harten, A., Ebben, M.J.R., 2002b. Waiting times at periodically switched one-way trac
lanes: a periodic two-queue polling system with random setup times. Probability in the Engineering and
Informational Sciences 15, 495517.
van der Zee, D.J., van Harten, A., Schuur, P.C., 1997. Dynamic job assignment heuristics for multi-server batch
operationsa cost based approach. International Journal of Production Research 35, 30633093.
Webster, F.W., 1958. Trac signal settings. Road Research Technical paper 39, Her Majestys Stationary Oce,
London.
Webster, S., Baker, K.R., 1995. Scheduling groups of jobs on a single machine. Operations Research 43, 692703.