14-th IFAC Symposium on Control in Transportation Systems

14-th
May IFAC
IFAC Symposium
14-th18-20, on
2016. Istanbul,
Symposium Control
Control in
on Turkey in Transportation
Transportation Systems
Systems
14-th
May IFAC
18-20, Symposium
2016. on
Istanbul, Control
Turkey in Transportation Systems
May 18-20, 2016. Istanbul, Turkey
May 18-20, 2016. Istanbul, Turkey
Available online at www.sciencedirect.com

ScienceDirect
IFAC-PapersOnLine 49-3 (2016) 291–296
Traffic
Traffic Flow
Flow Model
Model Validation
Validation Using
Using
Traffic
Traffic Flow Model
Flow Model Validation
Validation Using
Using
METANET,
METANET, ADOL-C
ADOL-C and
and RPROP
RPROP
METANET, ADOL-C and
METANET, ADOL-C and RPROP RPROP
Adam Poole and Apostolos Kotsialos
Adam
Adam Poole and Apostolos Kotsialos
Adam Poole Poole and and Apostolos
Apostolos Kotsialos
Kotsialos
School of Engineering and Computing Sciences, Durham University,
School
School of
of Engineering and Computing Sciences, Durham University,
School of Engineering
Stockton
Stockton
Road, DH1
Engineering
Road, DH1
and
and3LE,Computing
Durham,
Computing
3LE, Durham,
Sciences,
UK (e-mail:
Sciences,
UK
Durham
Durham
(e-mail:
University,
{a.j.poole,
University,
{a.j.poole,
Stockton Road,
Stockton Road, DH1 3LE, Durham,
apostolos.kotsialos}@durham.ac.uk)
DH1 3LE, Durham, UK
UK (e-mail:
(e-mail: {a.j.poole,
{a.j.poole,
apostolos.kotsialos}@durham.ac.uk)
apostolos.kotsialos}@durham.ac.uk)
apostolos.kotsialos}@durham.ac.uk)
Abstract: Macroscopic traffic flow model calibration is an optimisation problem typically solved
Abstract:
Abstract:
by a derivative-freeMacroscopic
Macroscopic traffic
traffic flow
population flow
based model
model calibration
calibration
stochastic search is
is an
an optimisation
optimisation
methods. problem
problem
This paper typically
typically
reports on the solved
solved
use
Abstract:
by a Macroscopic
derivative-free traffic
population flow
based model calibration
stochastic search is an optimisation
methods. This problem
paper typically
reports on solved
the use
by
by a derivative-free
of aa gradient
derivative-free population
based population
algorithm using based
based stochastic
automatic
stochastic search methods.
differentiation.
search methods. The This
This paper
ADOL-C
paper reports
library on
reports on the
is coupled
the use
use
of
of aa gradient based algorithm using automatic differentiation. The ADOL-C library is coupled
with
of
with a gradient
the METANET
gradient
the METANET
based
based algorithm
source code
algorithm
source code
using
using andautomatic
and
this system
automatic
this system
differentiation.
is embedded within
differentiation.
is embedded
The
The ADOL-C
within ADOL-C
an
library
library is
an optimisation
optimisation is coupled
algorithm
coupled
algorithm
with
basedthe
with on METANET
the RPROP. The
METANET source
resultcode
source code and
and this
is a very system
efficient
this system is
is embedded
system which is able
embedded within
within to an
be optimisation
an calibrate METANET’s
optimisation algorithm
algorithm
based
secondon
based RPROP.
onorder
RPROP.model The
Thebyresult
result is
is aa
determining a very
verythe efficient
efficient
density system
system which is
is able
which equation
and speed able to be
be calibrate
to parameters
calibrate as METANET’s
METANET’s
well as the
based
second on RPROP.
order model Theby result is
determining very efficient
the density system
and which
speed is able
equation to be calibrate
parameters METANET’s
as well asextra
the
second
fundamental
second order model
model by
diagrams
order diagrams determining
used.
by used. Information
determining the
the density
obtained
density and speed
from
andfromspeed theequation
system’sparameters
equation as
as well
Jacobian provides
parameters well as the
asextra
the
fundamental
fundamental
insight into diagrams
the system used. Information
Information
dynamics. A 22 km obtained
obtained
site is from
considered the
the system’s
system’s
near Jacobian
Jacobian
Sheffield, UK provides
provides
and the extra
results
fundamental
insight into diagrams
the used.
system dynamics. Information
A 22process obtained
km siteare from
is considered the system’s Jacobian
near Sheffield, UK provides
UK and the extra
the results
insight
of a typical
insight into
into the system
system dynamics.
calibration
the and validation
dynamics. A
A 22 km
km site
22process is
is considered
siteare reported. near
considered near Sheffield,
Sheffield, UK and and the results
results
of
of aa typical
typical calibration
calibration and
and validation
validation process are reported.
reported.
of a typical calibration and validation process are reported.
© 2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Keywords: Road traffic, calibration, parameter estimation, optimization, gradient methods.
Keywords:
Keywords: Road Road traffic,
traffic, calibration,
calibration, parameter
parameter estimation,
estimation, optimization,
optimization, gradient gradient methods.
methods.
Keywords: Road traffic, calibration, parameter estimation, optimization, gradient methods.
1. INTRODUCTION FD are aggregate descriptions of the infrastructure-
1.
1. INTRODUCTION
INTRODUCTION FD
FD are
are aggregate
vehicle-driver system.descriptions
aggregate descriptions
The variation of
of inthe
the infrastructure-
infrastructure-
capacity and free
1. INTRODUCTION FD are
vehicle-driver aggregate
system. descriptions
The variation of inthe infrastructure-
capacity and free
In Poole and Kotsialos (2012) an optimisation formulation vehicle-driver vehicle-driver
speeds observed system.
in
system. realThe
Thedata variation
are
variation in capacity
projections
in capacity of and
the free
same
andsame
free
In
In
was Poole
Poole and
and
introduced Kotsialos
Kotsialos
for (2012)
(2012)
the an
an
macroscopic optimisation
optimisation
traffic formulation
formulation
flow model speeds
speeds observed
observed in
in real
real data
data are
are projections
projections of
of the
the same
In Poole and Kotsialos (2012) an optimisation formulation traffic flow
speeds adapting
observed in to local
real data inhomogeneities,
are projections e.g.
of thedrop
sameof
was introduced for the
the macroscopic
was solved traffic flow
flowofmodel
a ge- traffic
was
was introduced
calibration problem
introduced for which
for the macroscopic
macroscopic traffic
by means
traffic flow model
model lanes, flow
traffic
traffic flow adapting
adapting
or different
flow adapting
to
to local
traffic
to local inhomogeneities,
inhomogeneities,
composition.
local inhomogeneities, Variationse.g.
e.g. drop
e.g. drop
of that
drop
of
of
of
calibration problem The which was
was solved by
by means of
of aaa ge-
ge- lanes,
calibration
netic algorithm.
calibration problem
problem which
METANET
which was solved
Messmer
solved by means
and Papageor-
means of ge- systemor
lanes,
lanes, or
different
orshould
different
different
traffic
traffic composition.
be reflected
traffic on the FD Variations
composition.
composition. but all FDof
Variations
Variations
that
ofmodel
of that
that
netic
netic algorithm.
algorithm.
giou (1990); The
The
Kotsialos METANET
METANET
et al. (1998, Messmer
Messmer
2002) modeland
and Papageor-
Papageor-
was em- system system
system should
should be
be reflected
reflected on
on the
the FD
FD but
but all
all FD
FD model
model
netic algorithm. The METANET Messmer and Papageor- the same traffic
should beflow flow process.
reflected on the This FD does not
butnot mean
all FD model that
giou (1990); Kotsialos et
et al.
al. (1998, 2002)
box.model was
was em-em- the same
same traffic flow process.
giou
ployed,
giou (1990);
treated
(1990); Kotsialos
as a simulation
Kotsialos et al. (1998,
black
(1998, 2002)
2002) model
An additional
model was em- the optimisation
the same traffic algorithm
traffic flow process.
process. will This
This
equate
This
does
does
does not mean
all FD,
not mean that
that
since error
mean that
ployed,
ployed,
requirementtreated
treated wasas astheaa simulation
automaticblack
simulation black box. An
box.assignment,
spatial additional
An additional
additional i.e. thethe optimisation
optimisation algorithm
algorithm will equate
will equate
equate all FD,
all FD, since
FD, since error
since error
error
ployed, treated as a simulation black box. An minimisation
the optimisation is still the dominant
algorithm will objective.
all
requirement
requirement
determining was
was
the the
the
location automatic
automatic
and spatial
spatial
extension, assignment,
assignment,
of fundamental i.e.
i.e. minimisation
minimisation is
is still
still the
the dominant
dominant objective.
objective.
requirement the
determining was location
the automatic
and spatial assignment,
extension, of fundamental i.e. minimisation
The optimisation is stillproblem
the dominant objective.in Poole and
as formulated
determining
diagrams (FD). theThe location and extension,
motivation extension,
behind thisof ofis fundamental
fundamental
that current The optimisation problem as formulated
determining
diagrams
diagrams (FD).
(FD).
calibration(FD).
the
practiceThelocation
The either
and
motivation
motivation behind this
behindengineering
uses expert this is is that
is that current
that opinion The optimisation
Kotsialos
The optimisation
current Kotsialos (2012) is problem
a nonlinear
problem as formulated
mixed
as mixed
formulatedintegerin in
Poole
Poole and
inoptimisation
Poole and
and
diagrams
calibration practiceThe motivation
either uses behind
expert this
engineering current
opinion Kotsialos
problem. A(2012)
(2012)
geneticis
is a
a nonlinear
nonlinear
algorithm mixed
was usedinteger
integer
there optimisation
optimisation
in order to
calibration practice
to make a practice
calibration decision either either uses
about uses expert
the expert engineering
FD or engineering
use a separate opinion
opinion Kotsialos A
FD problem. (2012)
geneticis aalgorithm
nonlinear was mixed usedinteger
there optimisation
in order to
to
to make
make a decision
every aadiscrete
decision about
about the
the FD FD
FD or use
or use
usefroma separate FD problem.
demonstrate
problem. A
A genetic
the
genetic algorithm
soundness
algorithm of was
the
was used
approach.
used there
there in
Based
in order
on
order to
this
to
tofor make
for every decision road segment
about the resulting
or aa separate
separate
the model’s FD demonstrate the soundness of the approach. Based on this
FD demonstrate the soundness of the
the approach.
approach. Based on this
this
for every discrete
discrete
discretisation
for every discreterules. road
road
road Insegment
the firstresulting
segment
segment resulting from
from the
case, intuition,
resulting from
model’s
thepast
the model’s
model’s
work, a morethe
demonstrate
ex- work, aavariants
more
detailed
soundness
detailed
calibration
of
calibration
work using
work using
classic
Based
classicon and
and
discretisation
discretisation
perience, visual rules.
rules. In
In
inspection the
the first
first
and case,
case, intuition,
intuition,
preliminary data past
past
analysisex-
ex- work,
recent
work, a more
more detailed
of particle
detailed calibration
swarm
calibration work
work using
optimisation
using classic
(PSO)
classic and
and
discretisation
perience, visual rules. In the and
inspection firstpreliminary
case, intuition, data past ex- recent cuckoo variants
recent variants of
of particle
particle was swarm
swarm optimisation
in Poole(PSO)
optimisation (PSO) and
and
perience,
result to an
perience, visual
ad-hoc
visual inspection
approach
inspection andleading
and preliminary
away from
preliminary data analysis
data analysis
systems cuckoo
analysis recent search algorithms
variants
search of particle
algorithms swarm
was
reported
optimisation
reported in Poole
and Kot-
(PSO)
and and
Kot-
result
result to an
to an
that embed ad-hoc
an ad-hoc
ad-hoc
knowledge approach
approach
in their leading
leading away
away
own away
structurefrom
from and systems
systems cuckoo
sialos
cuckoo search
(2016).
search algorithms
These
algorithms was
evolutionary
was reported
algorithms
reported
the sialos (2016). These evolutionary algorithms were used for in
in Poole
were
Poole and
used
and Kot-
for
Kot-
result
that to
embed knowledge approach
in leading
their own structurefrom systems
and the sialos (2016).
calibrating theThese
Heathrowevolutionary
site used algorithms
in Poole were
and used
Kotsialosfor
that
display embed
of more knowledge
intelligent in their
forms own
ofown structure
automation, and
structure Kotsialos the sialos
and the calibrating (2016). theThese evolutionary algorithms were used for
that
display
display
embed
and Poole of moreknowledge
more intelligent
of (2013). intelligent
informs
In the latter
their of
forms automation,
automation, Kotsialos
of overparametrisation
case, Kotsialos calibrating
(2012) in addition
calibrating
is (2012) the Heathrow
the Heathrow
to a road
Heathrow
site
site
site
used
used in
stretch
used
in
in
near Poole
Poole and
Sheffield,
Poole and
Kotsialos
and Kotsialos
Kotsialos
which is
display
and Pooleof more
(2013). intelligent
In the forms
latter of
case, automation,
overparametrisation Kotsialos is (2012)
considered in
in addition
addition
here as to
to a
a
well.road
road
The stretch
stretch
results near
near Sheffield,
Sheffield,
reported which
which
demonstrate is
is
and
and Poole
a clear (2013).
risk(2013).
Poole In the
since typically
In the latter
latter
three case, overparametrisation
parameters
case, are necessary
overparametrisation is (2012) in addition
is considered here as to a road
well. The stretch
results near Sheffield,
reported which is
demonstrate
aaforclear
clear risk since
risk since
defining since typically three
typically three
a FD.typically parameters
three parameters
parameters are are necessary
are necessary considered
the validity
considered here
of
here as
the
as well.
well. The
approach.
The
necessary the validity of the approach. Optimal parameters were results
Optimal
results reported
reported demonstrate
parameters
demonstrate were
a clear
for risk the validity capturing
of the
the approach.
approach. Optimal parametersofwere were
for defining
defining aaa FD.
FD. determined
the validity of the essential Optimal characteristics
parameters the
The problem formulation suggested in Poole and Kotsialos determined
for defining FD. determined
underlying
determined
capturing
capturing
traffic
capturing
the
the
dynamics
the
essential
essential
as was
essential
characteristics
characteristics
shown in
characteristicsthe of
of the
the
ensuing
of the
The
The
(2012)problem
allowsformulation
problem formulation
the arbitrary suggested
suggested
selection in
in Poole
of FDand
Poole and Kotsialos
Kotsialos
location for underlying
underlying traffic
traffic seedynamics
dynamics as was
as was
was shown
shown(2016) in the
in the
thefor ensuing
ensuing
The problem formulation suggested in Poole and Kotsialos model validation,
underlying traffic Poole and
dynamics as Kotsialos
shown in more
ensuing
(2012) allows the arbitrary selection of FD location for model
(2012)
(2012) allows
homogeneous
allows the
road
the arbitrary
stretches,
arbitrary selection
which
selection of
of FD
themselves
FD location
are
location for
split
for details.validation, see Poole and Kotsialos (2016) for more
model
model
validation,
validation, see
see Poole
Poole and
and Kotsialos
Kotsialos (2016)
(2016) for
for more
more
homogeneous
into segments,road
homogeneous but stretches,
road stretches,
also penalised which
whichthe themselves
variance are
themselves are split
split details.
between details.
homogeneous
into segments, road
but stretches,
also which
penalised themselves
the variance are split
between details.
All the calibration methods used there are population
into
their segments,
parameters. but Thealso penalised
rationale the
behind variance
this between
penalisation
between All the calibration
into segments,
their
their parameters.
parameters.
is thatparameters.
butThe
by treating The The
alsorationale
the FD
penalised
rationale
the variance
behind
behind
as an extensive this
this penalisation
penalisation
quantity whose based All
based
All thetreating
the calibration
calibration the methods
methods
METANET
methods
used
used
there
there are
usedsimulator
there areaspopulation
are population
a simple
population
their
is that by treating the rationale
FD as an behind this
extensive penalisation
quantity whose based
executable treating
treatinginvokedthe
the METANET
METANET
for each fitnesssimulator
simulator
function as
as aa simple
simple
evaluation.
is that
start
is thatand by
and treating
end
by treating are the FD
decision
the as an extensive
variables
FD asvariables
an extensive in quantity
an whose
optimisation
quantity based
whose executable treatinginvokedthe METANET
for each fitnesssimulator
function as a simple
evaluation.
start
start
problem,and theend are
end parameter decision
are decision
decision variables
variance in
in an
penalty an
anwill optimisation
optimisation
result to This executable
This approach
executable invoked
follows
invoked forthefortheeach
commonfitness
each fitnesschoice function
choice made
function evaluation.
regarding
evaluation.
start
problem, and end
the are
parameter variables
variance in
penalty optimisation
will result to This approach
approach follows
the optimisation follows the common
algorithm common choice
used choice
for model made regarding
madeparameter
regarding
problem,
solutions
problem, the
that
the parameter
favour
parameter variance
similar
varianceFD. penalty
This kind
penalty will
of
will result
similarity
result to
to This
the approach
optimisation follows the
algorithm commonused for made
model regarding
parameter
solutions
solutions
was employed that
that favour
favour
as guidancesimilar
similarwhen FD.
FD. This
This
validatingkind
kind the of
of similarity
similarity
large the
the optimisation
estimation,
optimisationsee e.g. algorithm
Spiliopoulou
algorithm used
used
scale estimation, see e.g. Spiliopoulou et al. (2014) and Ngoduyet for
al.
for model
(2014)
model andparameter
Ngoduy
parameter
solutions
was employedthat as favour
guidancesimilarwhen FD.validating
This kindthe of large
similarity
scale estimation,
and Maher see see e.g. Spiliopoulou
(2012). Spiliopoulou
Here, a gradient et al.
al. (2014)
(2014) and Ngoduy
based optimisationNgoduy
was
model employed
of the as guidance
Amsterdam when
motorway validating
networks,the large scale
Kotsialos
scale and estimation,
Maher e.g.
(2012). et and
was employed
model
model of
of the
et al. (1998,
as guidance when
Amsterdam
the 2002).
Amsterdam motorway validating
networks,
motorwayconstraint
An additional
the large
networks,imposed Kotsialos
Kotsialosa method and
method
and Maher
Maher (2012). Here,
is introduced
(2012). Here, aa gradient
for solving
Here, gradient
a gradient
based
based optimisation
this calibration
based optimisation
problem.
optimisation
model
et al. of the
(1998, Amsterdam
2002). An motorway
additional networks,
constraint Kotsialos
imposed a method
Extra is
is introduced
introduced
information that for
for solving
solving
becomes this
this
availablecalibration
calibration
from the problem.
problem.
process
et
et al. (1998,
maximum
al. (1998, 2002).of An
number
2002). An additional
FD additional constraint
to be used constraint imposed
for a site. Itimposedis left up method
aa Extra is introduced
information that for solvingavailable
becomes this calibration
from the problem.
process
maximum
maximum number
number
to the optimisation of
of FD
FD to
to
algorithm be
be used
used
to for
for
decide a
a site.
site.
how It
It
many is
is left
left
FD upup
up Extra
of
Extra information
calculating
information partial that
that becomes
derivatives
becomes available
is highlighted
available
to of calculating partial derivatives is highlighted as well. The from
from the
as
the process
well. The
process
maximum
to the number ofalgorithm
optimisation FD to be to used for ahow
decide site.manyIt is left
FD to of calculating
gradient partial isderivatives
calculation derivatives
performedis is by
highlighted
use of theas asautomatic
well. The
The
to
be the
used optimisation
and over algorithm
which area toto decide
place how
them. many
decide how many FD to gradientFD to of calculating partial highlighted well.
to the
be
be used
optimisation
used andand over
over which
algorithmtotoplace
which areaarea to place them.
to place them. gradient calculation
gradient calculation is
calculation is performed
is performed by
performed by use
by use of
use of the
of the automatic
the automatic
automatic
be used and over which area them.
Copyright 2016 IFAC
2405-8963 © 2016, 291 Hosting by Elsevier Ltd. All rights reserved.
IFAC (International Federation of Automatic Control)
Copyright
Peer review©
Copyright © 2016
2016 IFAC
IFAC 291
291
Copyright ©under
2016 responsibility
IFAC of International Federation of Automatic
291Control.
10.1016/j.ifacol.2016.07.049
IFAC CTS 2016
292
May 18-20, 2016. Istanbul, Turkey Adam Poole et al. / IFAC-PapersOnLine 49-3 (2016) 291–296
differentiation algorithm ADOL-C, Walther and Griewank
(2012).
Section 2 provides a brief overview of the METANET
model. Section 3 outlines the optimisation problem formu-
lation. A brief site description and overview of available
data are given in section 4 and results are discussed in
section 5. Section 6 concludes this paper providing the
key areas of future work.
2. METANET MODEL OVERVIEW
METANET is a well known macroscopic traffic flow model.
A road network is represented as a directed graph con-
sisting of nodes and links. Links represent homogeneous
road sections, where the number of lanes is a constant
and there is no significant change of curvature or gradient.
Nodes are connected by links and are used at places where
the geometry of the motorway changes or at on-/off-ramp
junctions. Traffic enters via origin links and leaves through
destination links.
Time is discretised globally with a time step T and
the time horizon is K steps. Each motorway link m is
discretised into Nm segments of equal length Lm . The
variables describing traffic conditions in segment i of link
m, at time instant t = kT , k = 0, 1, . . . , K, are the traffic
density ρm,i (k) (veh/km/lane) of a link m with λm lanes,
the mean speed vm,i (k) (km/h) and the traffic flow qm,i (k)
(veh/h). The discrete time motorway second order traffic
flow model is the following.
T destinations, M4 the number of split junctions.
ρm,i (k + 1) = ρm,i (k) + [qm,i−1 (k) − qm,i (k)] (1)
L m λm
qm,i (k) = ρm,i (k)vm,i (k)λm
T diagram (4) equations. It includes the network-wide global
vm,i (k + 1) = vm,i (k) + {V [ρm,i (k)] − vm,i (k)}
τ parameters of the maximum density ρmax , minimum speed
T vmin and the mean speed equation (3) parameters τ , ν,
+ vm,i (k)[vm,i−1 (k) − vm,i (k)]
Lm
νT ρm,i+1 (k) − ρm,i (k)
− (3)
τ Lm ρm,i (k) + κ
where ν and κ are speed equation parameters and
V [ρm,i (k)] is the FD given by
  α 
1 ρm,i (k) m
V [ρm,i (k)] = vf,m · exp −
αm ρcr,m
where ρcr,m is the critical density of link m and αm a
parameter.
In order to account for speed drops due to on-ramp in-
flow the term −δT qµ (k)vm,1 (k)/ (Lm λm (ρm,1 (k) + κ))
is added at (3), where δ is a constant parameter, µ
the merging link and m is the leaving link. This term
is included only when the speed equation is applied to
the first segment of the downstream link m. Speed de-
creases due to weaving is accounted by adding the term
−φT ∆λρm,Nm (k)vm,Nm (k)2 /(Lm λm ρcr,m ) to (3), where
∆λ is the reduction in the number of lanes and φ is
another parameter. Constraints are imposed in the form
of a minimum speed vmin and a maximum density ρmax .
Traffic volume measurements at origins over the period
of K steps are required. Speed measurements can also
be used to better inform the model dynamics, but they
are not necessary. In order for the speed equation to be
applied at destinations s, measurements of the density
trajectories ρs (k) over the entire time horizon are provided
as boundary conditions as well. For a full description of the
METANET, see Messmer and Papageorgiou (1990) or its
manual, METANET (2008).
3. OPTIMISATION PROBLEM FORMULATION
Equations (1)–(4) applied on an arbitrary motorway net-
work can be expressed in the following discrete dynamic
state-space system form
x(k + 1) = f [x(k), d(k); z] . (5)
The state vector consists of the density and mean speed
of every link segment, i.e.
 T
x = ρ1,1 v1,1 . . . ρM1 ,NM1 vM1 ,NM1 (6)
where M1 is the number of motorway links in the network.
The disturbance vector d consists of the inflows qo entering
the system from entry points (origin links) like on-ramps
or the upstream main site boundary and optionally the
speeds vo at these locations; the densities ρs at the exit
locations (destination links) like off-ramps or downstream
main site boundaries; and the turning rates βnµ at every
split node n, where µ is the main out-link. Hence,
 µ T
d = q1 v1 . . . qM2 vM2 ρ1 . . . ρM3 β1µ1 . . . βMM4 4 (7)
where M2 is the number of origins, M3 the number of
z ∈ RΓ consists of the model parameters as encountered
(2) in the dynamic density (1), speed (3) and fundamental
φ, δ and κ. It also contains parameters related to the
fundamental diagram, i.e. vf , α, and ρcr .
A set of measurements y from a number of locations along
the motorway, are used for comparing reality and model
output. The resulting minimisation problem is
min J [x(k), y(k)] (8)
(4) z
subject to
x(k + 1) = f [x(k), d(k); z] , x(0) = x0 (9)
zmin ≤ z ≤ zmax (10)
where J [x(k), y(k)] is a suitable error function and zmin
and zmax are the lower and upper bounds, respectively, of
z’s elements. The evaluation of J at z requires the forward
integration of (9) given as input the measurements of x0
and d(k).
 the number of FDs used; each one’s parameters ρcr ,
Let N
α and vf are included in z, i.e.
 
 T . (11)
z = τ κ ν ρmax vmin δ φ vf1 α1 ρ1cr . . . vfN αN ρN
cr
 = 1 a single fundamental diagram is used. If
When N

N = M1 then every link has its own FD; this is the case for
292
 IFAC CTS 2016
May 18-20, 2016. Istanbul, Turkey Adam Poole et al. / IFAC-PapersOnLine 49-3 (2016) 291–296 293

the gradient based optimisation method presented here in where Zγ is the fundamental diagram index decision
order to avoid integer decision variables whose derivatives variable zγ refers to and wv,ρ,α is given from
are not defined. 
 w if zγ corresponds to a free speed
 v
Assume there are M5 loop detectors at the site giving wρ if zγ corresponds to a critical density
wv,ρ,α = (21)
speed measurements yj,v (k), j = 1, . . . , M5 . The measure-  wα if zγ corresponds to an exponent

ments vector for period k y(k) has the form 0 otherwise.
T
y(k) = [y1,v (k) . . . yM5 ,v (k)] . (12) From the chain rule
T
Speed measurements are sufficient for the validation of a ∂Jv (z) ∂x(z) ∂Jv (z)
second order model by virtue of the conservation equation, = . (22)
∂z ∂z ∂x(z)
as pointed out in Spiliopoulou et al. (2014).
A global list is retained assigning each sensor to the In view of (6) vector ∂Jv (z)/∂x(z) has the form
 
corresponding motorway link it belongs to. For each model T T T
∂Jv (z) ∂Jv (z) ∂Jv (z)
time period k the measurement and the model outputs are = ... , k = 1, . . . , K (23)
compared assuming the same measurement value, since the ∂x(z) ∂x1 (k) ∂xM1 (k)
model sample time is smaller than the sensor’s.
where
Measurement location j’s contribution to speed square
error terms is given by  T
 2 ∂Jv (z) T ∂Jv (z) ∂Jv (z) ∂Jv (z) ∂Jv (z)
= ...
Ej,v [x(k), y(k)] = yj,v (k) − vmj ,ij (k) . (13) ∂xm (k) ∂ρm,1 (k) ∂vm,1 (k) ∂ρm,Nm (k) ∂vm,Nm (k)
m = 1, . . . , M1 (24)
The total error is given by
K M5 Because of the quadratic error terms (13) and the fact
1  that Jv in (14) does not explicitly depend on the densities
Jv [x(k), y(k)] = Ej,v [x(k), y(k)] . (14)
KM5 j=1k=1
∂Jv (z)
In order to implicitly achieve the automatic assignment = 0 and (25)
∂ρm,i (k)
of similar FDs a penalty term Jp (z) is included in the
objective function. ∂Jv (z) 2
= [vm,i (k) − yj,v (k)] Im,i (26)
∂vm,i (k) KM5

N
 −1  
N
  2 ∀ m = 1, . . . M1 , i = 1, . . . , Nm , k = 1, . . . , K
Jp (z) = wv vf − vfr with Im,i a binary indicator function showing if there is
=1 r=+1 a measurement for segment (m, i) used in the error calcu-
 2  2  lation and j is the corresponding measurement station in
+wρ ρcr − ρrcr + wα α − αr (15)
(12).
where wv , wρ and wα are weighting parameters set to
0.001, 0.0015 and 1.0, respectively. Equations (23)–(26) allow for the analytical calculation of
one factor of the right hand side of eqn. (22); in order to
The problem’s objective function (8) takes the form complete this calculation and determine its left hand side,
J [x(k), y(k), z] = Jv [x(k), y(k)] + wp Jp (z) (16) the Jacobian matrix
 T T

where wp a weighting parameter, which depends on the ∂x(1) ∂x(1)
 ∂z ... 
problem size; here, wp = 5.0. ∂x(z)  1 ∂zΓ 

= ... ... ...  (27)
∂z 
By iteratively substituting (9) into (8) the optimization  ∂x(K) T ∂x(K)
T 
problem can be expressed as ...
∂z1 ∂zΓ
min J(z) (17) needs to be calculated, where
z
subject to T

∂x(k) ∂ρ1,1 (k) ∂v1,1 (k) ∂ρ1,N1 (k) ∂v1,N1 (k)
zmin ≤ z ≤ zmax (18) = ... ...
∂zγ ∂zγ ∂zγ ∂zγ ∂zγ
where the dependence on y(k) has been dropped because T
∂ρM1 ,NM (k) ∂vM1 ,NM (k)
these are measurements that don’t change. Gradient based 1 1
(28)
iterative methods for solving this problem require the ∂zγ ∂zγ
calculation of ∂J/∂z at each iteration. From (16) and because of (25) only the ∂vm,i (k)/∂zγ for every
∂J ∂Jv (z) ∂Jp (z) segment i of link m at time instant k with respect to every
= + . (19)
∂z ∂z ∂z model parameter zγ need to be calculated. It is exactly this
quantity that the ADOL-C AD library calculates at every
∂Jp (z)/∂z can be calculated analytically from (15) since simulation time step. Hence, in order to obtain ∂J(z)/∂z,
it consists of quadratic penalty terms and therefore ∂Jp /∂z and ∂Jv /∂x are calculated analytically from (20)

N and (25)–(26), respectively, and ∂x/∂z is calculated by
∂Jp (z)
=2 wv,ρ,α (zγ − z ) , γ = 1, . . . , Γ (20) ADOL-C during a simulation run with METANET con-
∂zγ figured with z. Having obtained this way ∂J(z)/∂z, a
=1,=Zγ

293
IFAC CTS 2016
294
May 18-20, 2016. Istanbul, Turkey Adam Poole et al. / IFAC-PapersOnLine 49-3 (2016) 291–296

Table 1. Traffic flow model parameters upper

and lower limits.
Variable τ κ ν vmin ρmax
Max. 60 90 90 8 190
Min. 1 5 1 5 160
Variable δ φ αm vf,m ρcr,m
Max. 4 3.0 5.00 130 40.0
Min. 0.001 0.00005 0.40 80 18.0
(a) Site sketch.
Table 2. Fitness function for three runs.
J Jv w p Jp
1st
repeat 1 48.50 39.66 8.84
repeat 2 49.21 36.96 12.25
repeat 3 51.37 40.47 10.90
8th
repeat 1 43.27 36.96 6.31
repeat 2 42.91 36.55 6.36
repeat 3 43.36 36.53 6.83
15th
repeat 1 52.17 33.73 18.44
repeat 2 52.37 32.13 20.24
(b) METANET site representation. repeat 3 62.21 38.92 23.29

Fig. 1. Sheffield site model (not on scale). Table 3. Optimal network-wide parameter sets
τ κ ν vmin ρmax δ φ
multistart version of the RPROP iterative optimisation 1st 22.67 22.40 56.68 6.18 182.75 0.08 0.00005
algorithm is used in the way described in Kotsialos (2013, 8th 11.39 29.17 41.96 8.00 173.30 0.00008 0.00005
2014) based on Riedmiller and Braun (1993). 15th 18.57 24.71 40.34 8.00 177.83 0.093 0.00005

A simple multistart initialisation scheme with points sam-

pled from a Latin hypercube is sufficient.

4. SITE DESCRIPTION AND DATA USED

The test site is the Northbound M1 motorway as it enters

Sheffield and can be seen in Fig. 1. It extends over 21.9 km
and the METANET model consists of 20 links. Recurrent
congestion has the form of a shock wave originating at the
centre. Usually it occurs at the end of the link 6 where
the off-ramp of Junction 33 is short and unable to cope
with the demand of exiting flow. Data collected by the
MIDAS system from Monday the 1st , 8th and 15th of June
2009 were used. They consist of flow, speed, and occupancy
measurements per lane averaged over one minute intervals.
It is well known that vm,i in eqn. (2) is the space mean
speed, which for a small area centred around a loop detec-
tor is estimated by the harmonic mean of individual vehicle
speeds passing over the detector. The loop detectors pro-
vide arithmetic mean speeds, but assuming homogeneous
traffic conditions along each lane, the harmonic and the
arithmetic lane mean speeds are the same. The cross lane Fig. 2. Optimal FD parameters and spatial assignment.
mean speed is estimated as the harmonic mean of the lane
speed measurements,. accounting this way for the lateral corresponding error component Jv and weighted penalty
speed variance but not the longitudinal. w p Jp .
Table 2 shows that the RPROP algorithm performs con-
5. RESULTS sistently for the three days and is able to find parameter
sets that result to an absolute average error on the order
5.1 Calibration of 6% to 7% of the speed for each detector station. Table
3 shows the network wide part of z from the best repeat
of the three days, whereas Figure 2 depicts each link’s FD
For the calibration optimisation problem solved, the upper parameters over space.
and lower bounds (10) are given by Table 1. For each day of
data the optimisation algorithm was repeated three times. Figure 3 is a heatmap distance-time diagram of the speed
Table 2 gives the values of total fitness function J and the based on data of the 15th. On the left the speed mea-

294
 IFAC CTS 2016
May 18-20, 2016. Istanbul, Turkey Adam Poole et al. / IFAC-PapersOnLine 49-3 (2016) 291–296 295

(a)

Fig. 3. Speed distance-time diagram for calibration results

based on the data of the 15th.

(b)

(c)

Fig. 5. Speed sensitivity over space and time with respect

to the FD parameters of links (a) L1 (b) L6 (c) L10.
Table 4. Validation Jv for the solution yielding
the best calibrated Jv on Table 2.
Data from the:
Optimal parameter 1st 8th 15th 8th
set from the: (incident)
1st 36.96 141.00 106.65 55.66
8th 338.26 36.53 327.48 103.38
Fig. 4. Distance-time of the speed with respect to the speed 15th 83.22 140.47 32.13 61.65
equation parameters for the data of the 15th.
this information for the benefit of planning systematic
surements used are shown and on the right the calibrated interventions to the traffic flow system.
model’s output. The model is able to represent traffic
dynamics and predict the congestion’s spatio-temporal 5.2 Validation
extension.
One of the additional benefits of calculating ∂J/∂z is the The quality of the best solution obtained by the calibra-
calculation of the Jacobian matrix ∂x/∂z. This provides tion process when data of a particular day were used is
the sensitivity of the speed at every instant and at every tested using the data of the other two days. The optimal
segment with respect to the model parameters. Figure parameter set of that day is given to METANET but
4 shows this sensitivity, ∂vm,i (k)/∂zγ , γ the appropriate the initial and boundary conditions are from the data of
index for z, for the speed equation parameters distributed the other two. The model’s output is then compared with
over space and time for the data of the 15th. Figure 5 the corresponding measurements from the site and Jv is
displays the sensitivity of the whole network’s speed over calculated.
space and time with respect to the FD parameters of links Table 4 provides the values of Jv when METANET uses
L1 at the start of the site, L6 at the middle, and L10 at the optimal parameter set z∗day (Table 3 and Figure 2)
the end.
that yields the best calibrated Jv in Table 2 (in bold).
Forwards and backwards waves are observed indicating Notice that this is not always the solution with the best
the dominant factors affecting speed at particular point in total fitness function value J. The diagonal elements are
time and space. These provide extra information regarding the calibration results and they are always smaller than
the flow dynamics which can be used for examining the the other elements of the matrix, as expected. It can be
influence of control measures or help target the most seen that z∗15th generalises better on the data of the 1st
influential areas where improvements need to take place. compared to z∗1st on the data of the 15th. Hence, from
Part of the future work is finding ways of exploiting these two days z∗15th is preferred.

295
IFAC CTS 2016
296
May 18-20, 2016. Istanbul, Turkey Adam Poole et al. / IFAC-PapersOnLine 49-3 (2016) 291–296
(a) data from the 1st
(b) data from the 8th with incident
Fig. 6. METANET simulation using z∗15th .
z∗8th does not generalise well on the data of the 1st
and 15th. A closer inspection of the measured speed
trajectories, lead to the conclusion that on that particular
day there was a sustained spill-back of congestion from
destination D33. Spillbacks are particularly difficult to be
captured by a macroscopic traffic flow model and distort
the calibration results. However, the optimal parameter
sets should remain valid even in the presence of their
effects. In order to verify this, an incident was introduced
for the duration of the spillback as indicated by the data
at link L6, just upstream of the exit to D33. The last
column of Table 4 shows the impact of introducing it. Jv
for both z∗1st and z∗15th has dramatically improved. Hence,
the calibrated models of the 1st and 15th remain valid for
the data of the 8th. This is visualised from the distance-
time diagrams shown in Figure 6. The results shown were
obtained by using z∗15th for the data of the 1st and 8th
with incident.
6. CONCLUSIONS AND FUTURE WORK
This paper has presented the combination of three sepa-
rate methods for modelling road networks (METANET),
automatic differentiation (ADOL-C) and numerical op-
timisation (RPROP) for developing a system of traffic
flow model parameter estimation. The additional require-
ment of automatically selecting each FD’s location and
extension is not satisfied to the highest degree possible
but is implicitly considered in the optimisation problem
formulation. The resulting system is capable of delivering
model parameter sets that remain valid under different
conditions. The developed system makes use of the gradi-
ent which is calculated using the ADOL-C library. In the
process of this calculation, the model’s speed sensitivities
with respect to the model parameters are determined.
This calculation yields additional information that can be
exploited for improving interventions to the traffic system.
The RPROP algorithm has showed that it is capable of
296
solving this highly complicated and demanding problem.
Its simplicity of implementation has allowed the integra-
tion of the three different source codes into a single system.
A first line of research following this work will be concerned
with model validation of other second order macroscopic
traffic flow models that do not suffer from the isotropic
assumption. Further work will also focus on improving
the software implementation and RPROP’s speed of con-
vergence. Implementing integer programming methods for
the problem formulation where a maximum number of
used FD constraint is imposed is another important area
of work. A more systematic investigation on how to use
the information contained in the sensitivity distance-time
diagrams for control purposes will need to be considered
carefully. More detailed experiments need to be conducted
in order to associate the correlation between the capacity
as calculated by the calibration and the measured traffic
composition.
ACKNOWLEDGEMENTS
The authors would like to thank EPSRC for partially supporting this
work and the Highways Agency for providing the necessary traffic
data and related information.
REFERENCES
Kotsialos, A., Pavlis, Y., Middelham, F., Diakaki, C., Vardaka,
G., and Papageorgiou, M. (1998). Modelling of the large scale
motorway network around amsterdam. Preprints of the 8th IFAC
Symposium on Large Scale Systems, 2, 354–360.
Kotsialos, A. and Poole, A. (2013). Autonomic systems design for
ITS applications. In 2013 16th International IEEE Conference
on Intelligent Transportation Systems, 178–183.
Kotsialos, A. (2013). Non-smooth optimization based on resilient
backpropagation search for unconstrained and simply bounded
problems. Optimization Methods and Software, 28(6), 1282–1301.
Kotsialos, A. (2014). Nonlinear optimisation using directional step
lengths based on RPROP. Optimization Letters, 8(4), 1401–1415.
Kotsialos, A., Papageorgiou, M., Diakaki, C., Pavlis, Y., and Mid-
delham, F. (2002). Traffic flow modeling of large-scale motorway
networks using the macroscopic modeling tool METANET. IEEE
Trans. Intell. Transp. Syst., 3(4), 282–292.
Messmer, A. and Papageorgiou, M. (1990). METANET: A macro-
scopic simulation program for motorway networks. Traffic Engi-
neering and Control, 31, 466–470; 549.
METANET (2008). ”METANET Documentation”. Dynamic Sys-
tems and Simulation Laboratory, Technical University of Crete,
Chania, Crete, Greece.
Ngoduy, D. and Maher, M. (2012). Calibration of second order traffic
models using continuous cross entropy method. Transportation
Research Part C, 24, 102–121.
Poole, A. and Kotsialos, A. (2016). Swarm intelligence algorithms
for macroscopic traffic flow model validation with automatic
assignment of fundamental diagrams. Applied Soft Computing,
38, 134–150.
Poole, A. and Kotsialos, A. (2012). METANET model validation
using a genetic algorithm. In Proc. of the 13th IFAC Symp. on
Control in Transportation Systems, 7–12.
Riedmiller, M. and Braun, H. (1993). A direct adaptive method for
faster backpropagation learning: The rprop algorithm. In Neural
Networks, 1993., IEEE International Conference on, 586–591.
IEEE.
Spiliopoulou, A., Kontorinaki, M., Papageorgiou, M., and Kopelias,
P. (2014). Macroscopic traffic flow model validation at congested
freeway off-ramp areas. Transportation Research Part C, 41, 18–
29.
Walther, A. and Griewank, A. (2012). Getting started with ADOL-
C. Combinatorial Scientific Computing, 181–202.