See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330741596

Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

Conference Paper · October 2015

CITATIONS READS

0 969

3 authors, including:

Carlos Aydos
Ohmio Automotion
9 PUBLICATIONS 46 CITATIONS

SEE PROFILE

All content following this page was uploaded by Carlos Aydos on 30 January 2019.

The user has requested enhancement of the downloaded file.

22nd ITS World Congress, Bordeaux, France, 5–9 October 2015

Paper number ITS-1644

Traffic signal splits optimisation in real time:

The Simplex algorithm in SCATS

J. Carlos Aydos*, Steven Shaw, Andrew Mehaffey

Roads and Maritime Services of NSW
PO Box 1927, Strawberry Hills NSW 2012, Australia
T. +61 2 93199420
[email protected]

Abstract
The objective of adaptive traffic signal control systems is to optimise traffic flow. This is
typically achieved by optimising intersection cycle length, phase splits and cycle offsets. In
real-time systems, a compromise between optimal and computable solutions is often achieved.
Phase split optimisation for trivial intersection problems are solved via efficient and ordinary
optimisation techniques. Realistic and practical intersections contain phases with overlapping
movements which require more sophisticated constrained optimisation techniques. This paper
presents a novel application of the Simplex algorithm that increases the number of cases for
which an optimal solution for phase splits can be found while maintaining its real-time
computability. This strategy increases the number of phases that can be optimised in SCATS
and enables traffic operators to choose the algorithm responsiveness.

Keywords:
SCATS, traffic signals, optimisation

Introduction

The Sydney Coordinated Adaptive Traffic System or SCATS (Sims & Dobinson 1980; Lowrie
1982) optimises cycle length, phase splits and cycle offsets. This paper focuses on an
improvement for the existing phase split optimisation strategy.

In typical signalised intersections, cycle length is split into two or more phases (a.k.a. stages or
patterns in some countries). Each phase corresponds to a time interval during which one or
Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

many intersection approaches are displayed a green light. The split optimisation is the process
by which the cycle length is split into optimal time intervals according to some objective.

In SCATS, this objective is a condition called equisat, or equal degrees of saturation (DS) for
every critical phase of the intersection. The traditional DS is defined by a volume-capacity ratio
(HCM-chap16 2000). In SCATS, DS is determined in a different manner to account for true
green time utilisation. The details of the DS definition are not relevant to the contributions of
this paper and can be found in Bliss et al. (2009). For the purposes of this paper, sufficies to
consider that DS is a traffic measure dependant on the phase green time interval.

The existing strategy used in SCATS is called incremental split plan selection or ISS and
provides optimal results under the assumption that traffic predicted for the next cycle is the
same as on the previous cycle. SCATS analyses the DS values of every phase in real-time and
adjusts the phase splits accordingly in order to achieve a predicted equisat condition on the next
cycle. In practice, the phase splits adjustments are purposefully dampened to change only a
fraction of the true desired change. The dampening avoids unnecessary large changes that can
be caused by traffic noise or traffic sensor noise.

ISS is optimal, but has two operational limitations: it can optimise a maximum of four phases
and the damping amount cannot be tuned by traffic operators. These limitations exist because
the current strategy executes an exhaustive search over the search space and an exhaustive
search becomes computationally challenging in real-time for larger number of phases or
dampening scenarios.

For example, let’s assume splits are calculated in integer percentage increments for a 2 phase
intersection. The search space contains all the splits of phases A and B which sum to 100%. For
example (A=0%, B=100%) or (A=1%, B=99%) or (A=2%, B=98%) and so forth. There are 101
possible solutions. Note that this computational analysis ignores practical constraints that splits
normally have a minimum size constraint (e.g. 5 seconds). Extrapolating the same analysis
above for 3 phases, the number of possible solutions results in 10201. Circa 1,000,000 for four
phases and 100,000,000 for five phases. It can be easily shown that the number generalises to
100^(n-1) where n is the number of phases.

The contributions of this paper are three novel strategies presented in the next 3 sections. One of
the strategies is the most applicable strategy and explained in detail with examples and results.

2
Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

Analytical Strategy

In traditional phase-based traffic signal control, the green interval of a signal group is contained
within a single phase. Such phase-based systems allow a simple and efficient optimal analytical
strategy. For example, Fig. 1 depicts the signal group schematics of a 3 phase intersection. The
arrows depict the movement directions controlled by the corresponding signal groups. For this
intersection, the optimum equisat splits can be calculated analytically. A procedure for
calculating equisat splits for pre-timed control can be found in the Highway Capacity Manual
(HCM-chap16 2000).

A phase B phase C phase

Figure 1 – Signal group schematic of a 3-phase intersection.

This section proposes an alternative strategy that is practical for a real-time control system
where new splits are calculated at every cycle similar to ISS. The optimal split s’A for A phase
depicted in Fig. 1 can be calculated as a percentage of the cycle length such as:

s’A= dsA . sA / (dsA . sA + dsB . sB + dC . sC) (1)

where si is the previous split for phase i during which dsi was the measured DS. The splits for B
and C can be calculated analogously. The derivation of the above equation is not shown here.
The above example is for 3 phases, but can be trivially extended to other number of phases.

The solution provided by Eq. 1 is optimal. However, on hybrid (phase and signal group-based)
control systems such as SCATS, signal groups’ green intervals can operate over multiple
phases.

One such example is depicted in the intersection signal group schematics in Fig. 2. This simple,
yet more realistic SCATS intersection has 3 phases and 5 signal groups. There is a signal group
(blue) that controls a movement continuously over phases A and B. This is called an overlap
(HCM-chap16 2000). This intersection has phases A, B, C and an overlap A/B. Generally
speaking, intersections can have multiple overlaps and overlaps can span more than two phases.
3
Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

A phase B phase C phase

Figure 2 – Signal group schematic of a 3-phase intersection with

1 overlap (blue signal group)

Each signal group has its own traffic sensor and own DS measurements, including the A/B
overlap. As a result, this intersection needs to optimise the green interval of A/B as well. This
poses an issue since A/B = A + B is a constraint. Optimal splits that generate a predicted equisat
condition in Eq. 1 do not necessary comply with this constraint. The analytical strategy does not
work even if a fourth equation is added to this example. Nevertheless, this is the preferred
strategy for intersections without overlaps.

Critical Path Strategy

Given the limited applicability of the previous strategy, this section expands the use of Eq. 1
into a new strategy that uses the concept of a critical path to manage constraints. This strategy
is easiest explained by depicting the intersection time intervals in a graph where edges are the
signal group intervals. See Fig. 3. The left most and the right most nodes are subsequent
instances of the same point in the cycle. One can observe there are two possible paths on this
graph: path A-B-C and path A/B-C. The basis of the strategy is to identify the critical path via
analysing DS for each signal group.

1 CYCLE

A/B
A B C

Possible Paths A-B-C, A/B-C

Figure 3 – Graph representation of the signal group intervals of
a 3-phase intersection with 1 overlap

4
Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

Akcelik (1981) demonstrates a similar critical path strategy for pre-timed control. This section
proposes an alternative strategy that is practical for real-time control. For the purposes of this
paper, the critical path is the path that has the highest lowest DS. In other words, find the lowest
DS of each path than find the path with the highest such DS. The critical path is than solved
using the strategy from Eq. 1. This is possible because a single path does not contain
overlapping intervals. For example, let’s assume the critical path for a given cycle was A/B-C.
In this case, the strategy of Eq. 1 would be applied for A/B and for C and s’A/B and s’C would be
determined as percentages of cycle length. Subsequently the alternative path can be recursively
solved, i.e applying the strategy of Eq. 1 to determine s’A and s’B as percentages of s’A/B.

A slightly more sophisticated example is depicted in Figs. 4 and 5 which show a 3-phase
intersection with 3 overlaps. This case has 4 possible paths. This results in 4 iterations of the
strategy in Eq. 1 instead of 2 iterations like in the previous example. Generalising, the number
of iterations equals the number of alternative paths.

A phase B phase C phase

Figure 4 – Signal group schematic of a 3-phase intersection with

3 overlaps

An unusual intersection design could exist such that there is no path within a single cycle. This
is depicted in Fig. 6. Solving this problem would require the use of the data over more than one
cycle. This is possible, however it raises questions about the applicability of this strategy.
Although this strategy is effective to typical intersections, the number of possible graph
topologies is very large and it is reasonable to assume that there could be a number of
unforeseeable special topologies to which this strategy would not be applicable.

Simplex-based Strategy

This strategy addresses the applicability issues of the previous 2 strategies. In 2 dimensional
search spaces, a generic hill climbing algorithm (Russell & Norvig 2010) searches values for X
and Y in an XY-plane that optimise an objective function by finding a local optima. That is the

5
 1 CYCLE

B/C C/A A/B B/C C/A A/B

B C A B C A B

Possible Paths A-B-C, A/B-C, A-B/C, B-C/A

Figure 5 – Graph representation of the signal group intervals of a 3-phase intersection with 3 overlaps

6
1 CYCLE

B/C C/A A/B B/C C/A A/B

Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

Figure 6 – Graph representation of the signal group intervals of a 3-phase intersection with 3 overlaps. Note this
example does not have individual phase intervals.
Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

X and Y values that produce the best Z. It does that via iterations. At each iteration, it moves X
and Y randomly and checks if the move improved the Z value. If positive, the algorithm adopts
the move as a successful move, otherwise rejects the move. If a large number of iterations is
performed, the sequence of moves converges to a solution. For convex objective functions, this
algorithm is guaranteed to converge to the optimal solution.

The Simplex algorithm (Press et al. 1996) is a specific type of hill climbing algorithm with a
more sophisticated logic. In 2 dimensional search spaces, a simplex is a triangle. In this case,
the Simplex algorithm is initialised by randomly choosing 3 vertices on this XY plane forming
a triangle. At each iteration, the vertices are moved systematically. Moves result in triangle
reflection, extension or contraction. Extensions result in a larger step size for the next move
while a contraction results in a smaller step size for the next move. This algorithm is equally
applicable for a larger number of parameters in higher dimensional spaces. Each iteration of the
algorithm can be summarized as follows:

// find the worst valued vertex

worst = select_worst(vertex1, vertex2, vertex 3)
// find the best valued vertex
best = select_best(vertex1, vertex2, vertex 3)
// find the reflected point by flipping the triangle over the side adjacent to worst
reflected = reflect(worst)
// find the extended point by double the distance of reflected
extended = extend(worst)
// find the contracted halfway point between worst and the side adjacent to worst
contracted = contract(worst)

if value(reflected) < value(worst)

if value(reflected) > value(best)
// replace worst with reflected
reflect
else
if value(reflected) > value(extended)
// replace worst with extended
extend
else
// replace worst with reflected
reflect
else
if value(contracted) < value(worst)
// replace worst with contracted
contract
else
// replace worst with contracted and contract the other two vertices as well
multiple contract

This algorithm is applied to find the equisat condition. For example, for a 3 phase intersection
with 1 overlap, only X= s’A and Y= s’B are parameters in the Simplex algorithm. In this case, s’C
and s’A/B are considered dependent variables such as

s’C = 100 - s’A - s’B (2)

s’A/B = s’A + s’B (3)

7
Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

The above formulation guarantees that the search space excludes solutions that violate the
constraints in Eq. 2 and Eq. 3. To complete the algorithm, an objective function needs to be
formulated that allows a vertex to be evaluated. This is calculated by taking the average
projected DS (ds’ave) of all phases and overlaps and subsequently calculating the squared error
against this average. For example, the projected DS of phase A is calculated:

ds’A = dsA sA / s’A (4)

For our 3 phase and 1 overlap example, the objective function is defined as:

(ds’ave – ds’A)2 + (ds’ave – ds’B)2 + (ds’ave – ds’C)2 + (ds’ave – ds’A/B)2 (5)

This function has a value of zero for a vertex at the equisat point and positive values that are
larger with increasing lack of optimality.

The above constraint formulation and selection of free parameters and objective function is the
main contribution of this paper.

70

60

50

40
sB

30

20

10

0
0 10 20 30 40 50 60 70
A
s
Figure 7 – Initialisation of the Simplex algorithm for a 3-phase
1-overlap intersection

8
Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

Fig. 7 depicts the starting point of the algorithm for this example. The horizontal axis
corresponds to the split for A as a percentage of the cycle and the vertical the split for B. The
XY plane is contour-coloured representing the Z- axis, i.e. the objective function in Eq. 5. The
dark blue part corresponds to the lowest part of the troth, i.e. the equisat condition. The
coloured part of this figure has a triangle shape. That is because the objective function is not
represented in areas of XY plane where traffic operation is disallowed. For example, A has a
minimum green limit of 5% (white area at the bottom), B has a minimum green limit of 5%
(white area at the left) and C has a minimum green limit of 5% (diagonal white area at
top-right). Remember, C is not explicitly modelled, but is a result of A and B via Eq. 2. The
triangle at the bottom represented by black lines contains the initial 3 vertices of the Simplex
algorithm. These initial vertices are chosen at random. Each vertex is represented by a colour
coded circle. The blue circle is the best valued vertex and the red the worst. The triangle needs
to move to a better part of the function. As a result, the vertex that is moved is the worst one.
The ‘x’ marks on the top of the dashed line, represent in sequence: the contraction point
candidate, the mid-point of the adjacent side, the reflection point candidate and the extension
point candidate. It is reasonably clear from this picture that the best candidate for a move is the
extension point, i.e. it lies at the edge of a colour closest to the deepest part of the troth. This
paragraph describes one iteration or move of the algorithm. Subsequent moves are performed
until a given stop criteria has been reached.

50
SPLITS %

40

30

20
t_0 t_1

150
A B C AB
DS

100

50
t_0 t_1
Figure 8 – Results of the Simplex algorithm for 3-phase 1-
overlap intersection

Fig. 8 shows the results of the application of the algorithm. Randomly selected DS values and
initial splits are used as an input to the algorithm. The vertical slices t0 and t1 represent the initial
and final states respectively. The top plot shows the splits and the bottom shows the DS values.
The results at t1 show that the DS values are not equalised. This is however, the optimal
9
Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

solution, i.e. the closest possible to equisat where Eq. 3 holds. Fig. 9 shows the same solution
when the A/B overlap is removed from the intersection. The figure shows that an equisat
condition is met.

Fig. 10 shows the algorithm successfully applied to 12 phases. So far, all the results were shown
without dampening, i.e. one shot solutions. As mentioned earlier, one shot solutions can lead to
excessively large step changes in splits that react to noise rather than traffic demand. Fig. 11
depicts a more practical numerical simulation of the algorithm over 10 traffic signal cycles
where a dampening factor of 0.5 is used. That is, only half of the change is applied to the split in
the optimal direction in the XY plane. For this simulation, the projected splits are used as inputs
for the next iteration. One can observe the convergence of the DS values to the equisat
condition.
SPLITS %

40

30

20
t_0 t_1

150
A B C
DS

100

50
t_0 t_1
Figure 9 – Results of the Simplex algorithm for 3-phase
intersection

10
SPLITS %

8
6
4
2
t_0 t_1

100

80
DS

60

40

t_0 t_1
Figure 10 – Results of the Simplex algorithm for 12-phase
intersection

10
Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

10

SPLITS %
8
6
4
2
0 2 4 6 8 10

100

80
DS

60

40

0 2 4 6 8 10
Cycles
Figure 11 – Results of the Simplex algorithm for 12-phase
intersection with a dampening factor of 0.5

Conclusion

This paper presented 3 novel strategies for the real-time optimisation of phase splits. The
analytical and critical path strategies both have application limitations. The section on the
Simplex based strategy provided details of the Simplex algorithm, a formulation of its objective
function and a formulation of traffic signal constraints into Simplex parameters. The
Simplex-based strategy is the most suitable strategy for providing greater operational flexibility
for SCATS. This strategy replaces the exhaustive search with an efficient sequence of
systematic steps that exploit the shape of the objective function. As a result, it allows (a) larger
number of phases to be optimised and (b) it enables traffic operators to choose different levels
of responsiveness of the algorithm via the choice of different values of the dampening factor.

Acknowledgements

The authors would like to thank the Roads and Maritime Services, New South Wales for
allowing the publication of this paper. Any opinions expressed or implied in this paper are their
own and do not necessarily reflect official organisation policy.

References

Akcelik, R., 1981. ARR 123 - Traffic signals: capacity and timing analysis.

Bliss, J., Currie, G. & Freeman, P., 2009. Guide to Traffic Management. Part 9: Traffic
Operations.

11
 Traffic signal splits optimisation in real time: The Simplex algorithm in SCATS

HCM-chap16, 2000. Signalised Intersections. In Highway Capacity Manual.

Lowrie, P., 1982. The Sydney Co-ordinated Adaptive Traffic System - principles,
methodology, algorithms. International Conference on Road Traffic Signalling (IEE
conference publication 207), pp.67–70.

Press, W.H. et al., 1996. Numerical Recipes in C, Citeseer.

Russell, S. & Norvig, P., 2010. Artificial Intelligence: A Modern Approach.

Sims, A.G. & Dobinson, K., 1980. The Sydney Coordinated Adaptive Traffic (SCAT) System
Philosophy and Benefits. Vehicular Technology, IEEE Transactions on, 29(2), pp.130–137.

12

View publication stats