J. Inst.

Maths Applies (1971) 8, 164-185

Delay-minimizing Settings for Fixed-time Traffic

Signals at a Single Road Junction

RICHARD E. ALLSOP

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

Research Group in Traffic Studies,
University College London, England

[Received 5 October 1970 and in revised form 31 December 1970]

The operation of traffic signals at road junctions is controlled wholly or partly by preset
schedules. Expressions for the delay at fixed-time traffic signals have previously been used
in several methods of calculating signal settings. These methods give satisfactory results
in many cases, but have a number of drawbacks which are largely overcome in a new
method described in this paper. This method finds settings that minimize the estimated
average delay per unit time to all traffic passing through the road junction, subject to cer-
tain constraints imposed by the traffic engineer. It first determines whether there are any
settings that both satisfy these constraints and allow all the approaching traffic to pass
through the junction in the long run. If there are, it is shown that delay-minimizing
settings exist and are essentially unique, and an iterative procedure is given which can be
shown to converge to the delay-minimizing settings. A computer program for carrying
out the procedure is briefly described, and a few illustrative results are given.

1. Introduction
TRAFFIC SIGNALS are commonly used to share the carriageway between conflicting
traffic streams at busy road junctions. There are two main types of control for traffic
signals. Fixed-time signals operate always to a preset schedule, whilst the operation
of vehicle-actuated signals is controlled partly by vehicles approaching the junction.
Like fixed-time signals, however, vehicle-actuated signals require a preset schedule,
to which they operate when there is heavy traffic on all roads approaching the junction,
or when faults develop in the vehicle-actuation apparatus.
A road leading to a signalized junction may be such that all the vehicles waiting
there to enter the junction form one queue; alternatively, it may be possible for the
vehicles to form two or more separate queues according to the directions they intend
to take at the junction. In the former case the road will be said to form one approach to
the junction, whilst in the latter case it comprises as many approaches as there are
separate queues. An approach has right of way when traffic from that approach is
allowed to enter the junction.
In order to define the delay to a vehicle passing through the junction from a certain
approach, suppose that there is a point on the approach so far from the signal that,
even if the signal is red, approaching vehicles have not yet begun to slow down when
they pass the point. Suppose also that there is another point beyond the junction
such that vehicles that have had to stop at the signal have ceased to accelerate when
164
 FIXED-TIME TRAFFIC SIGNALS 165

they pass this second point. Then the delay to a vehicle is the difference between the
time it takes to travel between these two points and the time it would take if the
approach concerned always had right of way and the junction were not obstructed by
other traffic.
The purpose of traffic signals is to regulate the times at which the various approaches
have right of way, and two approaches are said to be compatible if they can have right
of way simultaneously. Considerations of safety and convenience usually determine
whether any two approaches are compatible.

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

At a fixed-time signal, two or more sets of mutually compatible approaches have
right of way in turn in cyclic order for predetermined times. One repetition of this
process is called a signal cycle, and its duration is called the cycle-time. The part of
the cycle for which one set of approaches has right of way is called a stage.
Suppose that a junction has n approaches, numbered arbitrarily from 1 to n, and
that the signal cycle has m stages, numbered in the order of their occurrence in the
cycle. Then the order in which the approaches receive right of way can be specified
by an m x n matrix A = (a^), called the stage matrix, such that
_ J l , if approach j has right of way in stage i
'J \0, otherwise.
One restriction is usually imposed in practice on this matrix: if the same approach
has right of way in two or more stages, then these stages must be consecutive (stage m
and stage 1 being regarded as consecutive).
Consider now just one approach; the restriction just imposed on the stage matrix
means that the approach has right of way for just one period in each cycle. Thus the
signal controlling the approach will show red, red and amber, green, and amber once
each in turn during the cycle. It is found, however, that, when estimating the delay
to traffic, the cycle may be regarded as comprising simply a period of effective red
time followed by a period of effective green time (see, e.g. Webster, 1958). The effective
red time for an approach is the duration of the part of the cycle during which, when
estimating delay, it is assumed that no traffic passes the signal. The effective green time
is the remainder of the cycle-time, during which, when estimating delay, traffic is
assumed to pass the signal at a constant rate, called the saturation flow, so long as
there are vehicles waiting on the approach.
The saturation flow usually corresponds closely to the average flow observed on the
approach when the signal has been showing green for several seconds but there are
still vehicles waiting. Many practical aspects of the relation between the saturation
flow and the actual movement of traffic on the approach, and of the relation between
the effective green time and the time for which the signal actually shows green are
discussed by Webster & Cobbe (1966).
When estimating delay the traffic is regarded as consisting of identical vehicles
equivalent in size and performance to the average passenger car. Traffic flows are
expressed in terms of these passenger car units (p.c.u.), and actual vehicles of various
types are regarded as equivalent to different number of p.c.u.'s (Road Research
Laboratory, 1965).
The time that is effectively green for every approach that has right of way in a
particular stage in the signal cycle will be called the effective green time for that stage,
and the time from the end of the effective green time for one stage to the beginning
166 R. E. ALLSOP

of the effective green time for the next stage will be called the lost time following the
former stage. There are thus as many periods of lost time in the signal cycle as there
are stages. Some or all of each period of lost time can be effectively green for one or
more approaches, because an approach having right of way in two successive stages
usually has right of way also during the lost time between those stages, and the effective
green time for an approach may overlap either the lost time before the first stage in
which the approach has right of way, or the lost time following the last such stage,
or both. If such overlaps occur, their duration is called the extra effective green time
for the approach concerned.

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

Previous methods of calculating signal settings (Webster, 1958; Miller, 1963, 1964)
give satisfactory results in straightforward cases and have justly found widespread
application. They are, however, based on a number of approximations and take into
account delay to traffic on only one of the approaches having right of way in each
stage. This approach is the one on which the ratio of the arrival-rate to the saturation
flow is largest, and is called the representative approach for the stage. The previous
methods are also difficult to apply when there are approaches having right of way
in more than one stage of the signal cycle, and when the effective green times for the
stages are subject to specified minima.
This paper describes an improved method for calculating the cycle-time and the
effective green time for each stage in the signal cycle at a junction controlled by a
fixed-time traffic signal. These times are determined so as to minimize the estimated
average delay to all traffic passing through the junction. The data required are the
average arrival-rate and the saturation flow on each approach, and the amount of
lost time occurring after each stage in the signal cycle. The junction layout and the
sequence in which the approaches have right of way are subject only to the restriction
that each approach can have right of way for only one period in each cycle. This
period may comprise one or more stages. A maximum can be imposed on the effective
green time for each stage, and a maximum can be imposed on the cycle-time. The
method can also be used in cases where the green times for some of the stages, or the
cycle-time, or both, are specified.

2. Estimate of Delay
In order to choose delay-minimizing signal settings it is necessary to estimate the
average delay to all traffic passing through the junction as a function of the signal
settings. Various expressions for the average delay to one p.c.u. passing through a
fixed-time traffic signal from one approach have been derived by Clayton (1940),
Wardrop (1952), Beckmann, McGuire & Winsten (1956), Webster (1958), Meissl
(1962, 1963), Miller (1963), Darroch (1964), Newell (1965), and McNeil (\96Sa,b).
Webster's expression is used here because it has been tested by comparison with
observed and simulated data, it requires measurement of saturationflowsand average
arrival-rates only, and it is reasonably simple in form.
Of the other expressions, only Miller's has been similarly tested, and his expression
is less simple and requires measurement of the variance of the number of p.c.u.'s
arriving per cycle. NewelFs also requires this measurement, and the other expressions
that allow for the randomness of arrivals all require for their evaluation the location
of numerous zeros of functions of a complex variable.
 FIXED-TIME TRAFFIC SIGNALS 167

3. Mathematical Statement of the Problem

Let m = the number of stages in the signal cycle,
n = the number of approaches,
A = (au) be the m x n stage matrix,
c = the cycle-time,
If = the effective green time for stage /,
Lt = the lost time following stage /,
m

L = £ Li = the total lost time,

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

i= 1
Ao = L/c,
and X = (Xo, Xu A 2 , ..., Xm).
Then the signal settings are completely determined by the vector X. By definition of'
the lost times Lt and that of Ao,

i= 0
It will be convenient to write
S(X) = t A,.
»=o
Consider now the approaches. For approach j let
AjC = the effective green time
and aOj = the proportion of the total lost time that is effectively green.
Then
0^aOJ<l (j = 1,2, ...,«),
and, since aki = 1 or 0 according as approach j does or does not have right of way
in stage /,
m
AjC = a0JL+ £ a^Xfi,
i=l
or, since Ao = Lie,
m

i= 0

Again, for approach 7 let

dj = the average delay per p.c.u.,
qj = the average arrival-rate,
Sj = the saturation flow,
X
J= #A>
and )>j = ^ / ^ (jj- will be called the flow ratio for approach^).
Then dj is estimated (Webster, 1958) by
d =
< ro{£-AJl)+2j-x.}-
It will be convenient to rewrite this in the form
168 R. E. ALLSOP

and to define, for./ = 1, 2, ...,«,

and

Following Miller (1963), the total delay per unit time on all the approaches to the

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

junction will be called the rate of delay for the junction.
Since, on approach j , an average of q} p.c.u.'s per second are delayed by an average
of dj seconds, the rate of delay for the junction is

In the above expression for dj, the coefficient 9/10 is the same for all approaches,
so that the rate of delay for the junction is proportional to

J=l
The vector k will be chosen to minimize D subject to the constraint S(k) = 1, and to
certain other constraints which will now be discussed.
In order to ensure that, in the long run, every approach has right of way for long
enough to allow all arriving traffic to pass through the junction, it is necessary that
Xj < 1 0 = 1,2, ...,«),
i.e.
m

Z M i > ^ 0 = 1 , 2, ...,»).
i= 0
These constraints express in terms of X. the condition that, with the resulting signal
settings, the capacity of each approach must exceed its average arrival-rate. They will
be called the capacity constraints.
If a maximum is imposed upon the cycle-time, let it be cM, and if a cycle-time is
specified, let it be cs. Then the constraints c ^ cM or c = cs can be written

/•o > — or lo = - .
C C
M S
In any case, the cycle-time must be positive, so that, if neither of the above constraints
applies, then Xo > 0. (Strictly, Xo > 0, but it will be seen in Section 5 that as A= -> 0,
D -» oo, so that the X, that minimizes D cannot be affected by permitting, for the
present, Xo = 0.) Let
L\cM, if it is required that c < cM,
L\cs, if it is required that c = cs,
0, otherwise.
Then the cycle-time constraint can be written
= k0, if the cycle time is specified,
k0, otherwise.
 FIXED-TIME TRAFFIC SIGNALS 169

From the fact that the lost time can only be part of the cycle-time it follows that
0 < k0 < 1. .
If a minimum is imposed on the effective green time for stage /, let it be giM, and if
no such minimum is imposed let gm = 0. Then it is required that Xf ^ giM,
(/ = 1, 2, ..., m). Let

Then, since Xo = L\c, the minimum green time constraints can be written

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

Xi 3= M o 0 = 1' 2» •••> m).
If the cycle-time is specified, or is subject to a maximum, then this cannot be less
than the sum of the lost time and all the minimum green times. Hence the kt
(« = 1, 2, ..., m) must satisfy the condition

£ fc;< 1 - 1
k
if fco#O.
i=l O
It will later be required that fcf < 1 (1 = 1, 2, ..., m), and the following argument
shows that this implies no loss of generality. Suppose that the largest of the imposed
minimum green times exceeds the total lost time. Then let this minimum green time
be added both to the amount of lost time that is effectively green for each approach
having right of way in the corresponding stage, and to the total lost time. Finally,
let the largest minimum green time be replaced by zero. The effective green time for
each approach is unchanged, but none of the minimum green times now exceeds the
total lost time.
Another type of green time constraint that can be provided for is the specification
of effective green times for one or more stages. Any such green times are added both
to the total lost time and to the amounts of lost time that are effectively green for the
approaches having right of way in the relevant stages. The other green times and the
cycle-time are then determined as if the stages whose green times are specified did
not appear in the signal cycle. There can be at most m — 1 stages with specified green
times {in — 2 if the cycle time is specified), for otherwise the settings would be com-
pletely determined by the constraints.
A value of k satisfying all these constraints will be called a feasible solution, and a
feasible solution that minimizes D(X) will be called an optima! solution.
The problem of finding an optimal solution is discussed in Sections 4-6. Throughout
the discussion it will be useful to denote real space of/? dimensions by R" and the vector
in Rm+l whose rth component is 1, and all other components are zero by e,-
(; = 0, 1,2, ...,m).
EXAMPLE. Consider a crossroads controlled by signals with a two-stage cycle, traffic from
north and south having right of way in stage 1, and that from east and west in stage 2.
Suppose that the traffic forms just one queue on each of the four roads, so that there
are four approaches, and let these be numbered in the order N, S, E, and W. Then

0 1 1
and
170 R. E. ALLSOP

where
) 0 1 2
Suppose also that none of the lost time is effectively green for any approach except
number 4, for which the proportion is a04.. The the capacity constraints take the form

and

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

Suppose that a maximum is imposed on the cycle-time and minima on the green
times; these constraints are
AQ ^ KQ,

and A2 ^ k2A0.
Figure 1 shows the intersection of the plane A0+A1+A2 = 1 with the positive octant,
and the lines in this plane which correspond to the various constraints. The part of
the plane corresponding to the set of all feasible solutions is indicated.

The shaded area of the plane o , ^

together with the heavily ruled parts of its
boundary represents the set of all feasible
solutions
FIG. 1. Constraints and feasible solutions for a junction controlled by a signal with a two-stage
cycle subject to a maximum cycle-time and minimum green times.

4. Existence and Location of a Feasible Solution

The first-step-towards finding a value of X that minimizes -DQJ),-subject to all the
constraints, is to find some feasible solution; any such solution will serve as a starting
 FIXED-TIME TRAFFIC SIGNALS 171

point. The conditions imposed on the kt (i = 0, 1, 2, ..., m) in the last section ensure
that there exists at least one X satisfying simultaneously the cycle-time constraint, the
minimum green constraints, and the constraint S(X) = 1. Such a X will not necessarily
also satisfy the capacity constraints. If the flow ratios are too large, no feasible solution
exists; this corresponds to the junction being overloaded with traffic.
A necessary and sufficient condition for the existence of a feasible solution, and an
algorithm for finding one whenever one exists, will now be given. Consider the set of
values of X satisfying the following amended constraints, in which the capacity con-

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

straints have been relaxed so that they can be used in an application of the simplex
method of linear programming. The amended constraints are
m
£ a0Ai 5= y} (j = 1,2, ...,«),
i=O
/.Q = KQ or XQ ^ *o>
and Aj ^ kiX0 (i = 1, 2, ..., m).
LEMMA 1. If any X satisfying the amended constraints exists, then S(X) attains a minimum,
subject to the amended constraints.
Proof. If any X, satisfying the amended constraints exists, let one such X be Xt. Then,
since kx > 0 (i = 0, 1, 2, ..., m), the set of X satisfying the amended constraints and
such that S(X) < S ^ ) is a closed and bounded set in Rm+1. Being a continuous func-
tion of X, S(X) therefore attains a minimum on this set; let it do so at X = Xo. Then
S(X0) must also be its minimum subject to the amended constraints.
Let a0 = 1 — S(X0), with Xo defined as in the Lemma; then a0 is denned if and only
if there exists a X satisfying the amended constraints.
THEOREM 1. A feasible solution exists if and only ifa0 is defined and either (a) a0 = 0
m
andX0 does not lie in any ofthe hyperplanes Xo = 0 and £ a^Xi = yjU= 1.2, . . , « ) ,
or (b) aQ > 0.
Proof. If <70 is undefined, no X satisfying the amended constraints exists, and hence
no feasible solution exists.
If (To < 0> then S(X0) > 1, and, by definition of Xo, no feasible solution exists.
m
If (70 = 0 and Xo lies in one or more of the hyperplanes k0 = 0 and £ aiikl = yjt
then one or more of the Af must be increased to satisfy the capacity constraints or to
make the cycle-time finite. This would make S(X) > 1; hence no feasible solution
exists. If <T0 = 0 and Xo does not lie in any of these hyperplanes, then Xo is a feasible
solution.
If (T0 > 0 then

is a feasible solution according as the cycle-time constraint is Xo = k0 or Xo ^ k0,

since each of these values of X makes S(X) = 1, and the equal increases in the com-
ponents of X satisfy any capacity constraints not already satisfied by Xo and do not
violate the minimum green constraints because 0 < k, ^ 1 (i = 1, 2, ..., m).
The evaluation of Xo is a standard problem in linear programming, soluble by the
simplex method (Vajda, 1961).
172 R. E. ALLSOP

5. Existence and Uniqueness of Delay-minimizing Settings

If no feasible solution exists, then there are no signal settings satisfying all the
constraints discussed in Section 3, and in order to calculate settings, the possibility of
relaxing some of the constraints would have to be examined. On the assumption that
a feasible solution does exist, it will be shown in this section that an optimal solution,
and hence delay-minimizing settings, exists and that its uniqueness, or otherwise,
depends only on the stage matrix.
It is now necessary, as foreseen in Section 3, to restrict the set of feasible solutions

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

slightly by requiring that Xo > k0, instead of Xo ^ k0, if k0 = 0, i.e. if no maximum
is imposed on the cycle time.
The constraint S(X) = 1 restricts the feasible solutions to a subspace, isomorphic
to Rm, of Rm+1. If the cycle-time is specified, the feasible solutions are further restricted
to a subspace isomorphic to R1"'1. Within these subspaces, since the capacity con-
straints are strict inequalities, the feasible solutions do not form a closed set. The
existence of an optimal solution is established in Theorem 2 by showing that values of
X that minimize D(X) are confined to a closed subset of the set of feasible solutions.
A corollary confirms that even if the cycle-time is not subject to any maximum the
optimal solution corresponds to a finite cycle-time.
THEOREM 2. If a feasible solution exists, then so does an optimal solution.
Proof. For all feasible X, i0 is positive and, for j = 1, 2, ..., n, /}(Ay) > 0 and
<7,(Ay) -+ oo as Ay -» yJt so that D(X) -* oo as Ay -»y 3 . Now let ^ x be any one par-
ticular feasible solution. For each j there exists a Y} such that if ys < Ay < Yj then
D(X) > 0(1^. It follows that when X = Xu Ay > Yj (j = 1,2,..., ri).
In the cases where the cycle-time constraint is Ao = k0 or Ao ^ k0 > 0, consider
the closed sets defined, in the subspaces {S(X) = 1} n {Ao = &o} and S(X) = 1
respectively, by the minimum green constraints together with the inequalities Ay ^ Yj.
D(X) is continuous on these closed sets and is therefore bounded below and attains its
lower bound there; let it do so at Xo. Then for any X in either of these closed sets,
D(X) > D(k0).
In particular
D&.J > D(k0).
But, for any feasible solution X not in the appropriate closed set,
D(X) > DQ.J.
Hence D(X) ^ D(X0) for all feasible X, and Xo is an optimal solution.
It remains to consider the case where the cycle-time constraint takes the form
Ao > 0. It can be shown that D(X) -> oo as Xo -> 0 uniformly for all ?.u l2, ..., Am such
that X is feasible. Hence there exists a Yo > 0 such that, for all feasible X with Xo < Yo,
D(X) > DQ.J.
The proof for the case where the cycle-time constraint took the form A0 > k0 > 0
shows, on putting k0 = Yo, that an optimal solution k 0 with D(X0) ^ D(Xt) exists
subject to the constraint Ao ^ Yo. It has just been shown that, for all feasible A with
Xo < Yo, D(X) > D(Xj). Hence Xo remains optimal when the constraint l0 > Yo is
removed; i.e. an optimal solution exists when the cycle-time constraint takes the form
;.o > o.
 FIXED-TIME TRAFFIC SIGNALS 173

The uniqueness or otherwise of optimal solutions will now be investigated. The

constraint S(X) = 1, and possibly the cycle-time constraint, are linear equalities that
restrict the feasible solutions to a subspace of Rm+1. All the other constraints are linear
inequalities in the lt (i = 0, 1, 2, ..., ni). The set of feasible solutions is therefore
convex. This will be used to show that if more than one optimal solution exists, then
the line joining them has the direction of a vector p = (ji0, filt \i2, ..., nm) such that
Ho = 0 and (jit, \i2, ..-, nm) belongs to the null-space in Rm of the stage matrix A.
The optimal solution is therefore unique if A has rank m, and if not, any point between
two optimal solutions on the line joining them will be shown to be an optimal solution.

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

The interpretation of these results is that all optimal solutions correspond to the
same cycle-time and to the same effective green time on each approach; they are
therefore indistinguishable so far as the traffic is concerned. There can be more than
one optimal solution only if the stages are such that different allocations of green time
between the stages can result in the same effective green time on each approach.
If there is more than one optimal solution, it is sufficient to find one of them in order
to determine delay-minimizing settings.
If X is a feasible solution, then the directions of all vectors \i such that X+h\i is a
feasible solution for all sufficiently small positive h will be called the permissible
directions at X.
In matrix operations, all vectors will be regarded as row-vectors. Let
a = (floi>ao2> •••,flon)
and

The transpose of a matrix or vector will be denoted by the superscript T, and the gradi-
ent operator in Rm+1 by
) d d d

Let
A = (A!,A 2 , ...,A n ).
Then
XA* = (A0A).
Let
B(3i) = (b^X))
be the (m+1) x (m+1) matrix of second derivatives of D(X), and let the (n+1) x (n +1)
matrix B*(X) = (b?j(X)) be defined by

° 0 = 1,2, ...,n),

and
KM = b%\) = 0 (1 ^ i < j «S n).
 174 R. E. ALLSOP

Then B* is the matrix of second derivatives of D when D is regarded as a function of

the n +1 variables Xo, A ^ A 2 , ..., An.
The following four lemmas are required in the proof of Theorem 3, the main theorem
of this section. Detailed proofs of these lemmas are given elsewhere (Allsop, 1970).
LEMMA 2. Ifk is a feasible solution, then the function D satisfies the conditions of Taylor's
Theorem {Hardy, 1955) in the neighbourhood ofk; in particular, ifk+\i is also a feasible
solution, then

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

where 0 < h < 1.
LEMMA 3. B(*.) = A*B*(3i)A*r.
The follows from the fact that dAJd^i = au (/ = 0, 1, ..., m;j - 1, 2, ..., n).
LEMMA A.Ifk satisfies the capacity constraints and Ao > 0, the matrix B*(k) is positive
definite.
It is sufficient (Mirsky, 1955) to show that, with some permutation of the rows and
columns of B*(X.), all the leading minors are positive. It can be shown that this is
true when the order of the rows and columns is reversed.
LEMMA 5. A feasible solution "k0 is optimal if and only if, for all permissible directions
Ho at Xo, \i0 . V£>(X0) > 0.
THEOREM 3. Ifk0 andko+\i are two optimal solutions, then fi0 = 0 and(jiu fi2, •••, A O
belongs to the nullspace in Rm of A.
Proof. By Lemma 2, Taylor's Theorem can be applied to expand D(ko + fi) about k0,
giving

where 0 < h < 1.

But Xo and X 0 +n are both optimal, so that
D(X0+v) = D(k0),
and, using Lemma 3,
H . \D(ko) + ±(nA*)B*(ko + hn)(.\iA*y = 0.
Moreover, by Lemmas 4 and 5, neither term in this expression can be negative. Hence
each is zero, and, again by Lemma 4,
jiA* = 0.
i.e.
^o = 0,
and
m
£ Hfiij = 0 X = 1,2, ...,n).
i= 0
Hence

i.e. 0*!, fi2, ..., nm)A = 0 as required.

COROLLARY 1. If A has rank m, there is only one optimal solution.
COROLLARY 2. If there is more than one optimal solution, all such solutions correspond
to the same cycle-time and the same effective green times on all approaches.
 FIXED-TIME TRAFFIC SIGNALS 175

Proof. Let any two optimal solutions be Xo and then = >.0 + u, where u has
the properties proved in the theorem. Hence

but the components of these two vectors are the values of L/c and the Ay corresponding
to the two solutions, and the corollary follows.
6. An Algorithm for Finding Delay-minimizing Settings
In this section it will be shown how, starting from any feasible solution X,x (found,
for example, by the methods of Section 4) a sequence {kp; p = 1, 2, 3, ...} of feasible

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

solutions can be constructed, which converges to an optimal solution. It would in
principle be possible to choose X,p+1 so that the direction of Xp+1 — Xp is that permissible
direction at Xp in which the gradient of D(X) is least (i.e. to use a method of steepest
descent). To simplify calculation, however, \p+1 will (except in the special circum-
stances considered in Lemma 6) be defined in terms of X.p by an equation
V i = K+hp(ek-ei)>
where 0 < k, I ^ m if the cycle-time is unspecified, 1 < k, I ^ m if the cycle-time
is specified, k # /, and hp > 0. The subscripts k and /, which determine the direction
°f ^-p+i—^p> a n d hp, which determines the distance of X,p+1 from X,p, will be determined
by kp, D(k) and the minimum green and cycle-time constraints. The form hp(ek—e() is
chosen because it is the simplest form of X,p+1 — Xp that ensures that the constraint
S(X) = 1 is satisfied throughout the sequence {kp}.
The choice of k, I and hp will be discussed, and then the convergence of {kp} to an
optimal solution will be established. This optimal solution corresponds to delay-
minimizing signal settings.
The components of Xp will be denoted by Xpi (/ = 0, 1, 2, ..., m), and | p will denote
evaluation at X. of a function of X. E.g. -^j- denotes the value of 3Dj3k0 when

K
In order that Xp+l shall be feasible, k and / must be chosen so that ek — e, is permis-
sible at Xp. This places no restriction upon k and / unless /tp0 = k, or Ap, = ktXp0 for
one or more i (i = 1, 2, ..., m), or both. If kpQ = k0, then / # 0, and if, for any i,
Xpi = MPo> then k # 0 and / # i for any such i. If the cycle-time is specified, then
Apo = &o for a\lp, and neither k nor / can be zero. All these restrictions are embodied
in the following rules for the choice of k and /. Let k be chosen so that
r . [3D if Api > fcjAp0 [i = 1, 2, ..., m) and the cycle-time
mm —
3D _ J Omm \ ^ i is unspecified.
if Api = for some i {i = 1, 2, ..., m) or the
mm —-i pj cycle-time is specified.
Since m > 1, this rule always yields a value of k. There may be more than one value
satisfying the rule, and, if so, one of them is chosen arbitrarily. Let / be chosen so that
3D)
max if A
p0
3D
31•ip
if Ap0 = k0.
 176 R. E. ALLSOP

This rule can fail to yield a value of / only if Ap0 = ^o and Xpl = ktXp0 0 = 1, 2, ..., m).
This is the trivial case in which the maximum or specified cycle-time is just the sum
of the minimum green times and the lost time, so that Xp is the only feasible solution,
and is therefore optimal, and the problem is solved. In other cases, there may be
more than one value of/satisfying the rule, and, if so, one of them is chosen arbitrarily.
It will be assumed in what follows that the rules yield values of both k and /. The
following lemma shows that if k and / are so chosen and Xp is not itself an optimal
solution, then either hp can be chosen to make D(Xp+1) < D(Xp), where Xp+1 =

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

X p +A p (e p -e,) or this inequality can be made to hold by making a suitable alternative
definition of Xp+i.
LEMMA 6. If k and 1 are chosen by the above rules and %p is not optimal, then either
-O(^P+I) < -O(^p) for all sufficiently small positive hp or Xpi = ktXp0 for at least one
i (i = 1, 2, ..., m) and there is a permissible \i at Xp, such that \i. \D(kp) < 0, of the
form
H= ± [ e o + 1 ^ - 0 + I £,>,],
ieC ieC
where Cis the set {/; 1 < i < m and Xpi = &jAp0} a n d / i s not in C.
Proof. If Xp is not optimal then, by Lemma.5 there exists a permissible n at Xp such
that n • VD(X.P) < 0.
Suppose first that Xpi > ktXp0 for / = 1, 2, ..., m. Then it follows from the rules for
choosing k and / that
dD 3D
^ ^dX, P m<o

But n is permissible, so that £ M< = ~ Z /*•» moreover | i . VZ)(Xp) < 0. Hence

^l>0 MI<0

and it follows that D(kp+1) < D(kp) for all sufficiently small positive hp.
Now suppose that D(X,P+1) ^ DQP) for arbitrarily small positive hp; then it follows
from Lemma 2 that
(dDdD\\

Hence, by the first part of this proof, Xpi =fcjAp0for at least one i (i = 1,2,..., m).
Let the sets C and F be defined by
C = {i; 1 < i < m and Xpi = k-^}
and
F = {/; 1 < i < m and Xpi > k^}.
Then it can be shown that (dDjdXk)\p = (dD/dXi)\p, and if their common value is G, it
follows that (dDldX0)\p < Gif Xp0 i= ko,(dDldX,)\p ^ GforalliinC.andaZ)/^;/,, = G
for all i in F. Since Xp is not optimal, there is a ji, permissible at X,p, such that
]i. VDQ.P) < 0. Using the properties just obtained for the components of VD(kp) it
can be shown that if fi0 = 0> then \i. \DQ.p) > 0, which contradicts the fact that
H . VD.Q,P) < 0; hence //0 # 0. This shows, in particular, that if the cycle-time is
specified and "kp is not optimal, then D(X+pl) < D(\p) for all sufficiently small hp.
 FIXED-TIME TRAFFIC SIGNALS 177

Now, for / in C, let jif = ktn0, so that y.f < nt for all such /. Let n% = n0, l e t / b e
the least / in F, let n* = - / < * - Z/'*> a n d l e t Vt = ° f o r a11 o t h e r » m F- T h e n
ieC
f1* = (/**> /*?> •••> f*m) is permissible at >.p and, using the properties of the components
of \D(kp), it can be shown that ji* . VD(kp) < 0. On putting n0 = ± 1, ji* takes the
required form ± [ e o + X^fc.e, —(1 + Z&i) e /]'
ieC ieC
To take account of the second alternative in Lemma 6, the definition of X,p+1 will
be extended as follows. If

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

0 + I*,OI(e*-e,) . \D(lp)\ < |[eo+ , p
ieC IEC IEC
then

ieC IeC

where A^ is to be determined, C is as defined in the lemma,/is the least i not in C, and

the sign is chosen to make (^ p+ i — Xp) . VD(X,P) negative.
It will be assumed in what follows that Xp is not optimal.
The choice of hp is influenced by two considerations: given the choice of k and /,
D(ip+1) should be as much less than D(Xp) as possible, but hp must not be so large that
y.p+1 is outside the set of feasible solutions.
It will first be shown that if the direction of Xp+1 — Xp is ek — e,, the capacity con-
straints and the weakest possible form of cycle-time constraint (i.e. Ao > 0) together
define (except in one special case) an upper bound, Hp, say, for hp. It will then be
shown that there is just one value of h in the interval 0 < h < Hp that minimizes
D(kp+h(ek—e,)), and hp will be given this value provided that the resulting Xp+1 is
feasible. If not, and in the special case where Hp is undefined, hp will be given the
largest value that makes Xp+1 feasible.
All the results can be extended to cover the case in which the alternative definition
of lp+1 is used (Allsop, 1970).
Consider first the implications of the constraints Ao > OandA; > yj(J = 1,2,...,«);
k p satisfies these constraints and Xp+l is to be obtained from "kp by increasing Xpk
by hp and decreasing Xpl by hp. Since a(y ^ 0 (i = 0 , 1 , 2, ..., m;j = 1, 2, ..., n), none
of the above constraints can be violated by the increase in lpk, but, except in one
special case, at least some of them will be violated by the decrease in ApJ if hp is too
large. The upper bound Hp that is thus imposed upon hp is different according as
(a) k = 0,
(b) / = 0,
(c) k / 0 and / # 0,
or (d) the alternative definition of kp+i applies.
In case (a), A(p+1)0 is certainly positive, and
Ajlp+i = Aj\p+(a0J-a,j)hp (j = 1, 2, ..., ri).
To make A 7 | p+1 > ys it is therefore necessary that
(atJ-a0J)hp < Ay|p-^ (J = 1, 2, ..., ri).
Since 0 < aOj < 1> this will certainly be true for those,/ for which atJ = 0.
178 R. E. ALLSOP

It is therefore sufficient that

for those j for which a / ; = 1 .

Hence, if A: = 0,

Hp = min
j

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

aij=l

Expressions for Hp in the other three cases can be obtained similarly and are given
elsewhere (Allsop, 1970). The one exception is that in case (c), if there is noy for which
al} = 1 and akJ = 0, then Hp is undefined. In such a case, every approach having
right of way in stage / also has right of way in stage k. There must therefore be at least
one approach has right of way in stage k and not in stage /; any such approach will
benefit, and none suffer, if green time is transferred from stage / to stage k. The green
time for stage / should therefore be made as small as possible in such a case, and it
will be seen later in this section that the algorithm achieves this.
If Xp+i = >.p+hp(ek—e,), then, once k and / have been chosen, D(kp+1) depends only
on the choice of hp. Let
d(h) = DQ.p+h(ek-e,)).
The following lemma establishes certain properties of this function of h.
LEMMA 7. IfHp is defined, then d(h) has just one turning value in the interval 0 < h < Hp,
and this is a minimum. If Hp is undefined, then d(h) is decreasing throughout the interval
0 < h < Xpl.
These results follow from the signs of the derivatives d'(h) and d"(h) (Allsop, 1970).
Let
h = ("the h in (0, Hp) such that d'(h) = 0ifHp is defined
° ~ I A ' if Hp is undefined.
Then Xp+h(ek—e,) satisfies the capacity constraints when 0 < h < h0. It remains to
examine the minimum green and cycle-time constraints. If all these are satisfied with
h = h0, hp will be given the value h0; if not, hp will be obtained by subtracting from
h0 an amount just sufficient to make kp+1 satisfy these constraints. Once again, the
cases (a), (b), (c) and (d) require separate examination.
In case (a), k = 0 and any of the minimum green constraints may be violated. The
fth constraint is just satisfied by h = ho—h*, where
lPi~Sa(h0-h*) = kikp0+ho-h*) 0 = 1,2,..., m),
<5j( being a Kronecker delta; i.e.

Hence, if A: = 0,

{»•
= h 0 - max JO, max
Expressions can be obtained similarly in the other three cases (Allsop, 1970).
It was shown in Theorem 2 that feasible solutions k such that D(X) ^ DQ.^ are
confined to a closed subset, Fu say, of the set of feasible solutions. The subset Fy is
 FIXED-TIME TRAFFIC SIGNALS 179

the set of feasible X such that Ao ^ Yo and A ; 5= Yi (j = 1, 2, ..., n), where Yo and
the y ; are as defined in Theorem 2. Thus Ft is a closed and bounded set in Rm+1 and
the sequence {Xp; p = 1, 2, 3, ...} constructed as just described, being such that
D(Xp) < £>(?.!) for all p, is confined to the set F x . It therefore has at least one limit
point in Fx. It will be shown in Theorem 4 that any such limit point that does not lie
in any of the hyperplanes bounding the set of feasible solutions is an optimal solution.
This result can be extended by similar but more tedious arguments to limit points
lying in the bounding hyperplanes.

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

Let X* be a feasible solution that is not optimal and is such that A? > k^*
(j = 1,2 m) and either A* > k0 or the cycle-time is specified. It is first necessary
to define, in terms of X*, a new function D*(X) for all X, sufficiently near to X*, and to
establish the continuity of this function. Let I* denote the evaluation at X* of any
function of X.
Suppose that k* and /* are chosen at X* by the rules given at the beginning of this
section; then since A* > ktX* and either X% > k0 or the cycle-time is specified, it
follows, as shown in the proof of Lemma 6, that (dDjdXk*—3Z)/dAf.)U < 0- Let Zbe
the set of all pairs (r, s) such that 0 < r, s ^ m if the cycle-time is unspecified, \ <, r,
s < m if the cycle time is specified, {dDjdX,)]* = \dDldXk*)\^ and (3D/aA,)U =

Then, since X* is feasible, the components of \D are continuous functions of X at

X* and there is a neighbourhood Nt of X* in Rm+1 or the subspace {Ao = A:o} (according
as the cycle-time is unspecified or not), such that, for all X in Nlt Ao > A:o if the cycle-
time is unspecified, A; > AT.-AQ (I = 1, 2, ..., m), dDjdXr—dDldls < 0 for all (r, s) in Z,
and if k and / are chosen at X by the rules given earlier then (k, I) is in Z. If follows from
the sign of dD\dXr—dD\dXs that, for all X in Nlt numbers hrs0(X) and hrs(X) can be
defined for X, r and 5 just as h0 and Ap respectively were defined earlier for Xp, k and /.
Then for X in Nx let

and
2>«(X) = min {Drs(X)}.
(r,s)eZ
LEMMA 8. D*(X) is a continuous function ofX in N^
Proof. By definition, hrs0(X) is either A5 or the appropriate solution of the equation

| r-es)) = 0.

Hence, either trivially or by the implicit function theorem (Hardy, 1955), hrs0(X) is
a continuous function of X in Nx. Moreover, hrs(X) is obtained by subtracting from
hrs0(X) a continuous function of X and is therefore itself continuous. Hence Drs(X) is
continuous in Nt for each pair (r, s) in Z, and so, therefore, is D*(X).
THEOREM 4. If X* is a limit point of the sequence {Xp}, A* > ktX* (i = 1, 2, ..., rri) and
either X% > k0 or the cycle-time is specified, then X* is an optimal solution.
Proof. Suppose that X* is not optimal. Then X* can be used to define D*(X) in the
neighbourhood A^x of X* as just described, and, as in Lemma 6, D,£X*) > 0 for all
(r, s) in the finite set Z, and hence D*(X*) > 0. Let
D*(X*) = 5.
12
180 R. E. ALLSOP

By the continuity of D(X) there is a neighbourhood N2, say, of X* such that for all X
in N2,
D{X*)-\d < D(X) < D(X*)+i8.
Further, by the continuity of D*(X) there is a neighbourhood N3, say, of X* such that
N3 <= Ni and, for all X in N3,
D*(X) > \b.
Now since X* is a limit point of the sequence {kp}, there is a value q ofp such that

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

Xg is in N2 n N3, and therefore in Nt. Hence the k and / chosen at Xq by the rules of this
section are such that the pair {k, I) is in Z and
D(Xq+1) = D(Xq)-Dkl(Xq)
< D(Xq) -D*(Xq)
< D(
(since Xq is in N2 n N3), i.e.

But {D(XP)} is a decreasing sequence, so that, for all/? > #, D(XP) < D(X*)~i8 and
Xp lies outside the neighbourhood N2 of X*, which contradicts the fact that X* is a
limit point of {Xp}. Hence X* is optimal.
This theorem and the corresponding results for limit points lying in the hyperplanes
bounding the set of feasible solutions show that the sequence generated by the algo-
rithm described in this section must converge to a value of X that corresponds to delay-
minimizing settings.

7. Possible Effect of the Third Term in Webster's Expression

The use of the function D(X) as a measure of the rate of delay does not take full
account of the third term in Webster's expression (1958) for the average delay per
p.c.u. on an approach. It has, in effect, been assumed that, for feasible X, the total
contribution made by the third term to the rate of delay for the junction is proportional
to the contribution made by the first two terms. Let the contribution of the third term
be C(X). Then

The values of the components of VC(X) usually give some indication whether the
inclusion of the third term would be likely to affect the delay-minimizing solutions
appreciably. Consider the components of \D(X) when X is optimal; if £ and / are chosen
at X by the rules of Section 6 then, as in the proof of Lemma 6, it follows that
dDldkk—dD\dki = 0 for all pairs (k, I) such that both ek — e, and e,—ek are permissible
directions at X. In most cases there is at least one such pair. Suppose that there is;
then 2D/9A; has the same value for all values of i that appear in such pairs. The values
of dC/dXt for these i will not in general be equal, and the range of these values, com-
pared with the common value of the dDjd)H, is an indication of the extent to which
the inclusion of C(X) could affect the optimal solution, and hence the delay-minimizing
settings. This range has so far been found to be small.
 FIXED-TIME TRAFFIC SIGNALS 181

8. Computer Program
A Fortran program has been written for calculating delay-minimizing settings using
the algorithm described in Section 6. The following data are required:
number of stages and approaches;
maximum or specified cycle time, if any;
minimum green time for each stage;
lost time following each stage;
average arrival-rate -j

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

saturation flow > for each approach;
extra effective green time-'
the numbers of the stages in which each approach has right of way.
The stages are numbered in the order in which they occur in the cycle; the approaches
may be numbered in any order. If required, the saturationflowscan be estimated by the
program from the dimensions and characteristics of the approaches, using the methods
of Webster & Cobbe (1966). Extensive data-checking is carried out, and if no feas-
ible solution is found to exist, a warning is printed, together with an estimate of the
extent to which the junction is overloaded.
The results provide the cycle-time and effective green times corresponding to the
optimal solution found by the algorithm, together with a table of estimated average
delays and rates of delay on the approaches, and the total rate of delay for the junction.
When the calculation for one set of data is complete the program can be required
either to amend this set of data and make further calculations, or to proceed to a
completely new set of data.
The running time on the University College IBM 360/65 is a few seconds for each
optimal solution found. About 33,000 bytes (8250 words) of core store are required
on this machine.

9. A Few Illustrative Results

Two examples of the application of the technique just described to realistic situations
will now be given, and the results compared with those obtained by applying Webster's
method (1958). In the first example, in which Webster's method is easy to apply,
the two methods give very similar signal settings and almost the same estimated rate
of delay. In the second example, however, in which Webster's method is less easy to
apply, the present method gives a much shorter cycle-time and an estimated rate of
delay 18% lower than Webster's.
In both examples the junctions are crossroads, and their layouts, with approach
numbers and saturation flows, are shown in Fig. 2.
EXAMPLE 1. Suppose that approaches 1 and 2 have right of way in stage 1 and ap-
proaches 3 and 4 in stage 2, that the lost time after each stage isfiveseconds, and that
no minimum green times or maximum cycle-time are specified. Settings have been
calculated for two ranges of average arrival-rates.
In the first range of arrival-rates, for which results are given in Table l(a), the flow
ratio on the representative approach of stage 1 varies from 1/3 to 6/10, and the corres-
ponding figure for stage 2 is always half as great. Thus Webster's method allocates
twice as much green time to stage 1 as to stage 2. The sum of these two flow ratios
182 R. E. ALLSOP

©
Example 2
© S t -4800
Example 1
S5=1656
S,= 1B00
-2083
-©
S,-6503 S 6 =2348-

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

S,=1800
,=1

© © ©
FIG. 2. Layouts of two junctions with approach numbers circled and saturation flows in p.c.u./h.

TABLE l(a)

Comparison of signal settings given by the present method and

Webster's method for the first range of arrival-rates in Example 1

Ratio of arrival-rate
on other approach to
that on representative Sum of flow ratios on representative approaches
approach (same for
both stages) 0-5 0-6 0-7 0-8 0-9

(i) Cycle time (in seconds) given by Webster's method

400 500 66-7 1000 2000
(ii) Cycle time (in seconds) given by the present method
0-2 41-4 510 670 98-6 193-2
0-4 40-3 49-7 651 95-8 187-7
0-6 39-9 48-7 63-6 93-2 182-3
0-8 40-2 490 63-2 91-7 177-3
10 42-7 52-8 69-2 101-6 198-6
g r e e n time for stage 1 .
fiift Ratio ~ / the
— g i v e n uj present method
green time for stage 2
0-2 1-89 1-90 1-91 1-94 1-97
0-4 1-90 1-90 1-92 1-94 1-97
0-6 1-90 1-90 1-92 1-94 1-97
0-8 1-89 1-89 1-91 1-94 1-97
10 1-89 1-90 1-91 1-94 1-97
(iv) Estimated rate of delay (in p.c.u.'s) for settings given by the
present method, with that for the Webster settings in brackets
0-2 3-7 (3-6) 5-5 (5-5) 8-6 (8-7) 151 (15-3) 35-6 (360)
0-4 4 1 (41) 6-2 (6-2) 9-6 (9-6) 16-6 (16-8) 38-5 (390)
0-6 4-7 (4-7) 7-0 (70) 10-7 (10-8) 18-4(18-7) 42-1 (42-9)
0-8 5-5 (5-5) 8 1 (81) 12-4(12-5) 21-0(21-3) 47-2 (48-3)
10 6-5 (6-5) 9-9 (9-9) 15-7(15-8) 27-8 (28-1) 66-3 (671)
 FIXED-TIME TRAFFIC SIGNALS 183

varies across the Table from 0-5 to 0-9 and, in Webster's method, determines the cycle-
time; the results are shown in part (i) of the Table. The ratio of the arrival-rate on the
other approach to that on the representative approach is always taken to be the same
for both stages, and varies from 0-2 to 1. The value of this ratio does not affect the
signal setting given by Webster's method; it does affect the cycle-time given by the
present method, though not, in this case, the allocation of green time, as parts (ii) and
(iii) of the Table show. The estimated rates of delay corresponding to the Webster
settings and to those given by the present method are compared in part (iv) of the
Tables; the difference is usually negligible and never exceeds about 2%.

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

TABLE l(b)

Comparison of signal settings given by the present method and

Webster's method for the second range of arrival rates in Example 1
Webster setting: cycle time = 120 seconds; equal green times for stages 1 and 2.

Ratio of arrival rate

on other approach to Ratio of arrival rate on other approach to
that on representative that on representative approach; stage 1
approach; stage 2 0-2 0-4 0-6 08 10

(i) Cycle-time (in seconds) given by the present method

0-2 114-6 112-9 111-2 109-8 116-5
0-4 111-4 109-8 108-5 1151
0-6 108-3 1071 113-5
0-8 1060 112-2
10 117-9
.... „ . green time for stage 1 . . . . ,
(n) Ratio : ; given by the present method
green time for stage 2
0-2 100 100 100 101 103
0-4 1-00 100 101 103
0-6 1-00 100 103
0-8 100 103
10 100
(iii) Estimated rate of delay (in p.c.u.'s) for settings given by the
present method, with that for the Webster setting in brackets
0-2 20-7(20-8) 21-7(21-8) 22-9(23-1) 24-6(24-9) 29-1(29-5)
0-4 22-7(22-8) 23-9(24-1) 25-6(25-9) 30-0(30-6)
0-6 25-1(25-4) 26-8(27-2) 31-3(31-9)
0-8 28-5 (290) 331 (33-6)
10 38-3(38-3)

In the second range of arrival-rates, for which results are given in Table l(b) the
flow ratio on each of the respresentative approaches is kept fixed at 5/12, whilst the
ratio of the arrival-rate on the other approach to that on the representative is allowed
to vary from 0-2 to 1 for each stage separately. Webster's method gives the same
setting in every case: a cycle-time of 120 seconds and equal green times for the two
stages. The cycle-times and allocations of green time given by the present method are
184 R. E. ALLSOP

shown in parts (i) and (ii) of the Table; most of the cycle-times are appreciably less
than 120 seconds. The estimated rates of delay corresponding to the Webster setting
and to the settings given by the present method are compared in part (iii) of the Table;
once again, the difference never exceeds about 2%.
EXAMPLE 2. There are six approaches and the signal cycle has three stages. The stage
matrix is
/I 0 0 1 0 0\
(l 1 0 0 0 0 ] .

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

Vo o i o i 1/
The lost times are four seconds following stages 1 and 3, and three seconds following
stage 2. A minimum green time of seven seconds is specified for each stage.
When the arrival-rates on the six approaches are 2923, 0, 60, 389, 522 and 205 p.c.u./
hour, the results given by the present method and by Webster's are as shown in Table 2.
TABLE 2

Comparison of signal settings given by the present method

and Webster's method in Example 2

Present method Webster's method

Cycle-time in seconds 53-3 910

Effective green times in seconds
Stage 1 16-3 39-9
Stage 2 70 70
Stage 3 190 33-1
Estimated rate of delay in p.c.u.'s 18-6 22-8

10. Comments
These results, and other examples in which both the present method and Webster's
have been applied, suggest that Webster's method will usually give a very good approxi-
mation to the settings that minimize the estimated delay in cases where that method
is easy to apply. The settings given by the present method sometimes differ appreciably
from Webster's approximations, however, particularly in that the cycle-time is often
shorter. Although the difference in estimated delay is usually small, it is sometimes
substantial. Moreover, the incidence of exceptional delay occurring, for example, when
the junction is partially blocked by an unusual accumulation of turning vehicles, is
likely to be reduced by the shortening of the cycle, because such delay is usually
confined to one cycle. Webster's expression for delay, like the other expressions
mentioned in Section 2, takes no account of such exceptional delay. The reduction in
cycle-time may well also be very useful when signals are linked. The circumstances
in which the two methods give appreciably different settings are not yet clear; to
clarify them, more extensive comparisons and practical trials are required. Only one
limited practical trial has so far been reported (Taylor & Allsop, 1969). The present
method has the advantage that it is just as directly applicable to complicated junction-
layouts as to simple ones.
 FIXED-TIME TRAFFIC SIGNALS 185

The work described here is included in a thesis approved by the University of

London for the award of the PhD Degree. Most of it was carried out while the author
held a Research Fellowship from International Computers Ltd. Example 2 in Section 9
comprises the estimated evening peak hour traffic flows at a proposed junction in
the County Borough of Huddersfield, and is quoted by permission of the Borough
Engineer and Surveyor, Mr A. L. Percy.

REFERENCES

Downloaded from http://imamat.oxfordjournals.org/ at University of California, Santa Barbara on September 8, 2015

ALLSOP, R. E. 1970 Optimisation techniques for reducing delay to traffic in signalised road
networks. PhD. Thesis, University of London.
BECKMANN, M., MCGUIRE, C. B. & WINSTEN, C. B. 1956 Studies in the economics of trans-
portation. Yale University.
CLAYTON, A. J. H. 1940 /. Instn civ. Engrs 16 (7), 247-284.
DARROCH, J. N. 1964 Ann. Math. Statist. 35, 38-88.
HARDY, G. H. 1955 A course of pure mathematics, 10th ed. Cambridge University Press.
MCNEIL, D. R. 1968a /. appl. Prob. 5 (3), 624-635.
MCNEIL, D. R. 19686 The delay at a traffic signal. Dept. of Statistics Technical Report No. 94.
Baltimore: Johns Hopkins University.
MEISSL, P. 1962 Mathematik Tech. Win. 9 (1), 11-18; (2), 65-69.
MEISSL, P. 1963 Mathematik Tech. Win. 10 (1), 1-4; (2), 63-68.
MILLER, A. J. 1963 Opl Res. Q. 14 (4), 373-386.
MILLER, A. J. 1964Proc. Aust. RdRes. Bdl (1), 342-365.
MIRSKY, L. 1955 An introduction to linear algebra. Oxford: Clarendon Press.
NEWELL, G. F. 1965 SIAM Rev. 7, 223-239.
ROAD RESEARCH LABORATORY 1965 Research on road traffic. London: H.M.S.O;
TAYLOR, J. & ALLSOP, R. 1969 Survr 133 (4024), 38^0.
VAJDA, S. 1961 Mathematical programming. Addison-Wesley.
WARDROP, J. G. 1952 Proc. Instn civ. Engrs 1 (2), 325-362.
WEBSTER, F. V. 1958 Traffic signal settings. Road Research Technical Paper No. 39. London:
H.M.S.O.
WEBSTER, F. V. & COBBE, B. M. 1966 Traffic signals. Road Research Technical Paper No. 56.
London: H.M.S.O.