Copyright ~ IFAC Control in Transportation Systems,
Braunschweig, Germany, 2000
PROGRESSION OPTIMIZATION IN PRIORITY ARTERIAL SIGNAL
NETWORKS
Chronis Stamatiadis, Nathan H. Gartner
Dept. ofCivil and Environmental Engineering,
University ofMassachusetts, Lowell, MA 01854, U.S.A.
Abstract: Arterial progression methods, adapted for application to grid networks, use mixed
integer linear programming models for the maximization of the bandwidths. Due to the
number of discrete variables involved, these methods are computationally demanding and
inefficient when applied to large networks. This paper develops a fast heuristic procedure,
which can be applied to both uniform and variable bandwidth optimization. The procedure
uses the traffic characteristics of the network and involves an iterative decomposition into
priority arterial sub-networks. This facilitates the determination of the optimal values for the
integer variables and enables the application of bandwidth optimization methods to large-scale
networks. Copyright ID 2000 IFAC
Keywords: Traffic Control; Bandwidth Optimization; Progressions; Mixed Integer Linear
Programming; Heuristics.
1.
INTRODUCTION
arterial. The MULTIBAND model is also formulated
as a mixed-integer linear problem.
A network
version of the variable bandwidth formulation was
also introduced (Stamatiadis and Gartner, 1996). The
multi-band design has been shown to provide
significant benefits in terms of common performance
measures such as delays, number of stops and fuel
consumption over conventional uniform bandwidth
models.
However, the multi-band formulation
increases the size of the continuous variable set and
of the constraint set by 20-60010 compared with the
uniform band formulation; while, the size of the
integer variables set remains the same. This makes
the model more computationally demanding.
Arterial progression methods, which use the width of
green bands along the arterial as the objective of
optimization, are often used for traffic signal control.
Numerous computer programs were developed for
finding optimal signal timing plans based on
bandwidth maximization. There are two categories
of such programs: one uses uniform bandwidth
optimization, while the other variable bandwidth
optimization.
Two of the most versatile and
advanced programs in the first category are PASSER11 and MAXBAND, which provide offset, split, cycle
length and left turn phase sequence optimization on
single arterial streets. The MAXBAND formulation
(Little et al., 1981) includes several decision
variables that are integers and, therefore. mixedinteger linear programming (MILP) is used for
solving the problem. The model is applicable to
single arterials as well as to networks (Chang et al.,
1988).
The primary difficulty in solving these models is due
to the multidimensionality of the discrete variables of
the problems: the large number of discrete variables,
as well as their wide feasible ranges, make the
solution of the program computationally demanding.
This paper presents a network decomposition
procedure for more efficient solution of the variable
bandwidth network progression problem. The procedure is based on the traffic characteristics of the
network. The network is decomposed into priority
sub-networks which facilitates the determination of
the optimal values for the integer variables. This
improves dramatically the efficiency of the solution
and thus enables to handle large-scale networks.
similar to the ones found in many metropolitan areas.
The solution that is obtained is equal or nearly equal
A more elaborate approach was developed by Gartner
et al. (1990, 1991) in the MULTIBAND model,
which is an extension of MAXBAND. By incorporating into the model a traffic-dependent criterion,
MULTIBAND calculates individual bandwidths for
each directional link of the arterial while still
maintaining main street platoon progression. The
individual bands depend on the actual traffic volumes
that each link carries and the resulting signal-timing
plan is tailored to the varying traffic flows along the
555
Outbound
Inbound
• WI+IJ ,
·,•
~
Outbound '""iiI!i!i5i!i!i2i!i\
Inbound -
ar:oIi~_
I!.,+IJ
'
,
~
+Oll.(/+IJ)
. rt·t/J
:
..
i.
..
+(1)).(1+ IJ)
:
..;
'I+IJ
.
,
,~
, '
.,-:,,
:~
.
~.
fy
..
'u
Si}
.
Figure 1: Time-space diagram for unifonn bandwidth optimization
4>(lj}.(lcI)
to the globally optimal solution obtained by solving
the original model. The procedure can be applied to
both unifonn and variable bandwidth optimization
models. The speed and the quality of the iterative
decomposition procedure may allow on-line
implementation of the method in the future.
2.
)Jk(, ji ~
I:!.ij =
'ij;
DESCRIPTION OF THE MODEL
queue
arterial);
outbound (inbound) red time at Sij;
w,,(wu )=
interference
time
time
for
speed on link Lij (ft/sec).
In the case of unifonn bands, the objective function
has the following fonn:
The geometric relations for the unifonn bandwidth
model are shown in Figure 1. Consider a network
with m arterials and each arterial having nj signalized
intersections. Let SI] denote the i'h signal on the Jh
arterial of the network and LI] denote the i'h link
(between signals i and i+ 1) of the Jh arterial, with) =
1 ,..., m and i = 1 ,... , n]. All time variables are
defined in units of the cycle time. The following
variables are defined:
cycle time (sec);
C=
outbound (inbound) bandwidth on
bj ( b; ) =
'y(ru )=
clearance
advancement of outbound (inbound)
bandwidth at Sij to clear turning-in
traffic before arrival of main-street
platoon;
outbound (inbound) progression
The multi-band optimization model consists of
several blocks of constraints dealing with the
individual arterials of the network as well as a set of
network constraints that ensures that continuous
bands are being produced on all the intersecting
arterials. The basic unifonn bandwidth optimization
fonnulation is presented first.
variables,
) =
internode offsets, time
from the center of the outbound
(inbound) red at SI] to the center of
the outbound (inbound) red at SkI;
directional node phase shift, time
from center of Tu to nearest center of
Maximize
f.(b) +k)
.~)
(1)
;=1
where k) is the target ratio of inbound to outbound
bandwidth for arterial).
The directional interference constraints ensure that
the progression bands use only the green time and
they do not cross through the red time. Only one
such constraint is needed for each signal Sij and each
directional band:
(2.a)
from
(2.b)
right (left) side of red at S,) to left
(right) side of outbound (inbound)
green band;
travel time on link i of arterial) in the
The arterial loop constraints result from the fact that
all signals must be synchronized, i.e., that they
operate with a common cycle time. In Figure 1 it can
be seen that for each link L ,) the summation of the
outbound (inbound) direction;
556
internode offsets and directional node phase shifts is
an integer multiple of the cycle time as follows:
cI>(ij),(I+I,i)
+ ~(lj),i+
+ ~i+I'J
-
~lj
=Kij
(6)
(3)
An additional important decision capability that can
be added through the Mll..P formulation of the
problem is in identifying the optimal left-turn phase
sequence with respect to the through green at any
signal Sij. Up to four additional integer variables may
be added at each node to identify all possible left-turn
phase sequences.
where 1C;j is an integer variable. The same principle
of signal synchronization applies to closed loops of
the network consisting of more than 2 links, resulting
in the network loop constraints. For simplicity, we
drop the arterial index in the notation of nodes and
internode offsets and we define the intranode offset
OJijk as the time from the center of the red at S; for
traffic moving from S, to Sf' to the center of red in the
crossing direction at the same node for traffic moving
The network loop
from Sj to Sk (Figure 2).
constraints specify that the summation of internode
and intranode offsets around a loop of intersecting
arterials must be an integer multiple of the cycle
time:
cl>ij
+roljk
+cI>jk +ro jkl +cI>kI
+ro kll +cI>/i +ro/ij
= Il n
In the MULTlBAND model the width of the
directional bands may differ from link to link. The
bandwidth can be individually weighted with respect
to its contribution to the overall objective function.
The link specific bands generated by MULTlBAND
are symmetric about the centerline of the arterial
progression band. The geometry of MULTlBAND is
shown in Figure 3.
The bandwidths and the
interference variables are redefined as follows:
(4)
where Pn is the integer variable of the nth network
loop. The number of network loop constraints and
the choice of a fundamental set of loops are given by
Gartner (2).
hi} (
~
Wij )jiT~l(
)
=
outbound (inbound) bandwidth of
link i on arterial j; there is now one
specific band for each link Lij;
=
the time from right (left) side of red
at Sij to the centerline of the outbound (inbound)
green band; the reference point at each signal has
been moved from the edges to the centerline of the
band.
The objective function now has the form:
..
1
n·-I
_
Maximize "
-L
~ (<x lj . hlj -Ki...
h)
L. n- -1
lj
lj
j=1
where
Fig. 2: Closed loop of intersecting arterials
<XI)
J
aij
and
(8)
1=1
are the link specific weighting
coefficients for the outbound and
inbound directions respectively. The
following coefficients are used:
The cycle time C (sec) and the link specific
progression speeds V'j and vij are treated as decision
variables as well. This formulation introduces considerable flexibility in the calculation of the best
progression scheme. Each of these variables must be
constrained by upper and lower bounds as follows:
Cl' Cl =
lower and upper bounds on cycle
length;
(eij , j,), (elj' IIj) = lower and upper bounds on
outbound (inbound) speed
Vi} (
'I+JJ
tl2?mazwZl2?i2:
T,
ezzzzmzzzzmnt
vlj );
hy» =
lower and upper bounds on
change in outbound (inbound) speed
(gl)' hij)' (gij'
vij (vij).
p2??2unaZZd
The corresponding constraints will be:
Cl
$
C
$
C2
(5)
Fig. 3: Geometric relations for the variable
bandwidth optimization model
557
where. Vy (~
Sy (
• use more economical mathematical programming
codes;
(9)
and
• to optimize larger-scale networks;
)=
outbound (inbound) directional flow
• obtain more reliable convergence to optimal
solutions;
rate on link L,;;
sij ) = saturation flow rate outbound (inbound)
• analyze a larger number of alternatives at a much
reduced cost; and
directional volume on link 11;;;
an integer exponential power; the values
0, 1,2 and 4 were used
P=
• enable the use of codes in real time, e.g., in a
multilevel RT-TRACS system (Gartner et al.,
1995).
In the case of variable bandwidths the band must be
constrained from both sides so that neither edge of
the band crosses through the red time. For each
signal Sy and for each link specific directional band
there are two interference constraints, as follows:
wij+bij/25,1-rij and 0~2/jib+ w
(10.a)
wij +bij/25,1-r;j
and wij 0~2/j,b+
Branch-and-bound is essentially a strategy of "divide
and conquer". The idea is to partition the feasible
region into more manageable subdivisions and,
eventually, to fathom the entire tree of integer
solutions. It is advantageous to decompose mixedinteger programming problems into smaller subproblems, in order to reduce the number of integer
variables that have to be considered in each subproblem, and to restrict as much as possible the
allowable range of each variable. Several authors
have proposed solutions for the uniform bandwidth
problem along these lines, e.g., Mireault (1991),
Chaudhary et al. (1991) and Pillai et al. (1994).
(lO.b)
The same relationship must be valid at both ends of
the band, i.e., at signals Sy and S;+lJ'
The arterial loop constraints, and the network loop
constraints are not affected by the variable bandwidth
extension of the procedure and remain as in
expressions (3) and (4). The cycle time and the
progression speed, as well as, the left-turn phase
in
pattern
constraints
remain
unchanged
MULTffiAND.
The new procedure that is described in this paper
does not merely exploit the mathematical structure of
the mixed-integer problem, as most other approaches
do, but is primarily based on the traffic characteristics
of the network. As such, it does not require any
modification of the optimization code as would be
required by some of the other approaches. A priority
arterial sub-network consisting of an arterial tree is
selected from the original traffic network based on its
geometry and on the volume of traffic that each link
is carrying. The priority sub-network is optimized
first and the results are then used for the solution of
the entire network. Alternative sub-networks can be
selected in a heuristic procedure if further improvements are desired.
Most importantly, the new
decomposition process does not require relaxation of
any integer variables unlike previous approaches.
The intermediate solution obtained are each a feasible
optimal solution for the priority sub-network
considered. This feature, in combination with the
fact that the priority sub-network contains the bulk of
the traffic volumes in the network, enables the
achievement of faster and better solutions. Results
indicate that an optimal solution is obtained in the
majority of cases during the first iteration. The
decomposition procedure is described below.
It has been shown (Gartner et al., 1990, 1991,
Stamatiadis and Gartner, 1996, 1997) that the multiband approach offers substantial flexibility for the
design of progressions that result in significant
It
improvements in all performance measures.
produces optimal variable progression schemes
tailored to the traffic demand and capacity of each
individual road section along each arterial street,
while simultaneously it optimizes progressions on the
crossing arterials as well.
3.
THE NETWORK DECOMPOSITION
APPROACH
The MULTffiAND model uses mixed-integer linear
programming for determining the optimal solution.
The principal difficulty in solving mixed-integer
problems is the number of integer variables and their
range, i.e., the size of the integer feasible set.
Virtually all MILP codes use branch-and-bound
strategies for calculating the optimal values. More
efficient computational procedures will result in the
ability to obtain improved solutions and, ultimately,
will lead to improved performance of the traffic
network.
Computational procedures are just as
important as are more accurate traffic models,
especially in an era of increasing real time applications. In the case of progression optimization in
networks this enables the following:
Let P J be a progression optimization problem
obtained by considering only a sub-network Ni of the
original network N, and P2 be another progression
optimization problem obtained by setting a selected
subset of the arterial loop and the network loop
integer variables in the network problem to a specific
Then the heuristic network
set of values.
'i'iR
The performance of the heuristic solution was tested
on two networks: the first in downtown Ann-Arbor,
Michigan (3 x 5 grid), shown in Figure Sa, and the
second in downtown Memphis, Tennessee (4 x 4
The grids are not
grid), shown in figure 5b.
complete, as there are not signals at each of the
nodes. The shaded arterials indicate the priority
arterial sub-networks utilized in this analysis. The
results are shown in Table 2. It must be noted that
the values of the objective function of MAXBAND
and MULTIBAND are not directly comparable. All
the runs were executed on a 200MHz Pentium
computer. The procedure was applied to both the
uniform and the variable bandwidth models. For the
uniform bandwidth case the heuristic calculated an
optimal solution - there are multiple optimal
solutions - for both sample networks. In the case of
the multi-band model, the heuristic located the
optimal solution only for one of the test networks.
However, the execution times were significantly
reduced by as much as 1:263 compared to the
original times.
set of values.
Then the heuristic network
decomposition method, shown in Figure 4, is as
follows:
~
Step 1: Identify a new priority sub-network N, c N
y Step 2: Optimize P1 for N" and save the resulting
~
values of the arterial loop and network loop integer
variables ICy', f.Jn' for all the links and network
loops ofN,.
Step 3: Optimize P2 by setting the integer
variables calculated in step 2 (IC;; = ICy'. f.Jn =f.Jn' :
1<:,1'
~
~
f.Jn E P1).
Step 4: Calculate the value of the objective
function. If it is better than the previous solution,
save it.
Step 5: Stop if all priority sub-networks have been
considered; otherwise go back to step 1.
A priority sub-network is chosen so that it contains
only a "tree" of arterials, eliminating any network
loops. The arterials contained in the "tree" should
include the principal arterials of the network and can
be chosen based on the following criteria:
1. Choose the principal arterial of the network to be
the trunk of the tree and include only crossing
arterials in the sub-network;
2. The tree consists of the maximum number of
arterial two-way links without forming any
network loops.
The solution of both P 1 and P2 can be obtained very
quickly due to the reduced number of integer
variables. Table 1 shows the number of integer
variables in the original problem and in the two subproblems P 1 and P2 of the network decomposition
approach for an m x n closed grid network (m x n
intersections and m+ n arterials) .
(b)
Select Priority Tree Sub-Network N, :N, c N
Optimiz. N, and save K.j: (ij)
Optimiz.]I; by setting K.j: (ij)
E
E
N,
N,
No
Reoort optimal
Fig. 5: The (a) Ann-Arbor,Michigan and the (b)
Memphis, Tennessee networks.
FigA The network decomposition procedure
559
Table 1: Size of MILP problem for an m
grid network.
Var.
MAXBAND
b,b
2(m+l)
MULTIBAND
2(m(n-I)+
+n(m-!))
z
x
5. REFERENCES
Little, lD.C., Kelson, M.D., Gartner, N.H. (1981).
MAXBAND: A program for setting signals on
arteries and triangular networks. Transportation
Research Record 795, 40-46.
n closed
Heuriltk
-Step1
HeuristkStep2
2(mn-l)
2(2mn-m-n)
Chang E.,C-P, Cohen, S.L, Liu, C, Chaudhary, N.A.,
Messer, C. (1988). MAXBAND-86: A program
for optimizing left-turn phase sequence in multiarterial closed networks. Transportation Research
Record 1181,61-67.
Gartner, N.H., Assmann, S.F., Lasaga, F., Hou, D.L.
(1990). MULTffiAND - A variable-bandwidth
arterial progression scheme. Transportation
Research Record 1287,212-222.
I
w,w
2(2mn-m-n)
2(2mn-m-n)
2(mn-l)
2(2mn-m-n)
K
(2mn-m-n)
(2mn-m-n)
mn-I
mn-m-n+1
Jl
(m-l)(n-I)
(m-I)(n-l)
0
(m-I)(n-I)
Total
int.
3mn-2m-2n+1
3mn-2m-2n+ I
mn-I
2(mn-m-n)
Example Networks / No. of integers
4x6
53
53
23
30
3x7
44
44
20
24
Gartner, N.H., Assmann, S.F., Lasaga, F., Hou, D.L.,
(1991). A multi-band approach to arterial traffic
signal optimization. Transportation Research
Vol. 25B(I), 55-74.
Stamatiadis, C., Gartner, N.H., (1996). MULTIBAND-96: A program for variable bandwidth
progression optimization of multi-arterial traffic
networks. Transportation Research Record 1554,
9-17.
Table 2: Objective Function Values and Execution
Times of the original and heuristic solutions.
Network
Memphis, TN (8/17)
Model
MAXB.
Obj.
Value
3.4682
Enc.
MULl.
7.9381
Ann Arbor, MI (8/14)
MAXB.
MULl.
2.9381
4.8930
6,735sec
15,012sec
4,235sec
9,995sec
Obj.
Value
3.4682
(100"10)
7.9381
(100%)
2.9381
(100%)
37680
(77%)
EIec.
50sec
(11135)
57sec
(1/263)
44sec
(1/%)
51sec
(1/1%)
Time
Time
4.
Stamatiadis, C., Gartner, N.H., (1997), Concurrent
Progression Schemes in Multi-arterial Traffic
Networks.. IFACIIFIPIIFORS Symp., Chania,
Greece.
Gartner, N.H., Stamatiadis, c., Tarnoff, P.l (1995).
Development of advanced traffic signal control
strategies for IVHS: A multi-level design.
Transportation Research Record 1494,98-105.
Mireault, P., (1991). A branch-and-bound algorithm
for the traffic signal synchronization problem
with variable speed. TRSTAN I, Montreal.,
Canada.
Chaudhary, N.A., Pinnoi, A., Messer, c., (1991).
Proposed enhancements to MAXBAND-86
program. Transportation Research Record 1324,
98-104.
Pillai, R.S., Rathi, A.K., Cohen, S., (1994). A
restricted branch-and-bound approach for setting
the left turn phase sequence in signalized
networks. Transportation Research, Vol. 32B(5).
CONCLUSIONS
This paper describes a new network decomposition
procedure for the solution of the mixed integer linear
programming formulation of the multi-band network
progression problem. It is shown that the procedure
can yield considerable reductions in computational
effort without or with little compromise in the quality
of the results. Reductions range from 1:100 to 1:300
compared with the original formulation with the most
dramatic reductions occurring in the case of the more
of
variable
bandwidth
complex
problem
optimization. By achieving these results one can
obtain more easily optimal solutions for large-scale
networks, analyze a larger number of alternatives, as
well as implement this strategy in an on-line system.
Overall, more efficient computational procedures
result in the ability to obtain improved solutions and,
ultimately, lead to improved performance of the
traffic network. They are just as important as are
more accurate modeling techniques.
560