Network Equilibrium Models: Varied and Ambitious · Network Equilibrium Models: Varied and Ambitious Michael Florian Center for Research on Transportation University of Montreal

INFORMS, November 2005 1

Network Equilibrium Models:Varied and Ambitious

Michael Florian

Center for Research on Transportation

University of Montreal


The applications of network equilibriummodels are varied:

They range from simple to the very complex

- single mode, single class equilibrium assignment - multi-class equilibrium assignment - generalized cost on road and transit networks - equilibrium assignment on congested transit networks - path analyses - complex multi-modal equilibration - combined mode trips - dynamic network equilibrium models


The single mode single class networkequilibrium model

( )0

minav

aa A

s x dx∈∑ ∫

subject to , ,

0, ,

( , )

ci

i

k ik K

k i

a ak kk K

h g i I

h k K i I

v h a Aδ

∈

∈

= ∈

≥ ∈ ∈

= ∈

∑

∑

kh

Numerous algorithms have been developped for its solution;As is well known the arc flows , are unique, but the pathflows , are not unique.

av


Some selected applications

• The presentation includes examples of the applicationof various network equilibrium models carried outaround the world

• The first set of applications was carried out with staticmodels of increasing complexity both for road andtransit networks

• The second set of applications presents new resultswith a dynamic network equilibrium model


Some Straightforward Applications

• Madrid, Spain• Pretoria, South Africa• Auckland, New Zealand


Madrid- AM Peak Flows and Speeds


Regional Study of the Province ofGauteng, South Africa

Study carried out byVela VKE – Pretoria, South Africa,


Scenario Without New facility


Scenario with New facility


Scenario comparison


Planning Transport in Auckland

Study carried out by AucklandRegional Council, Auckland, NZ


AM Vehicle Flows, 2001

36%

59%

5%MotorwaysArterialsLocal

Complex Variable Demand Network EqulibriumModels

INRO

Demand Model

EquilibriumTrip Assignment

Impedance

Step sizeMechanism

• The “Step Size Mechanism” maytake different forms depending on theknowledge that one has of theunderlying model.

• Does the model have an equivalentconvex cost optimizationformulation?

• Can the model be formulated as avariational inequality?

• The model is very complicated andone carries out “ad-hoc” feedback bysome averaging scheme.


• Some well known variants of such models are the Combined Distribution-Assignment Model , Combined Distribution-Assignment-Mode ChoiceModel, Equilibrium Assignment Model with Variable Demand,...

• One can use adaptations of nonlinear programming algorithms to obtain thesolution of such models

• The “Step Size Mechanism” may be trivially stated to be the result of a linesearch on the objective function

where xk is a current solution and dk is a direction of descent

Complex Variable Demand Network EquilibriumModels:

Equivalent Convex Cost Optimization Formulations

INRO

1

0 1min ( ( )), ( )k k k k k k kF x d x x x d x

λλ λ+

≤ ≤+ − = + −


It is well known that models are with asymmetric cost functions, such asintersection delays, transit travel time depending on auto travel times in multi-mode models can be formulated as :

Network EquilibriumModels:Variational Inequality Formulations

* *( )( ) 0a a as v v v− ≥

subject to , ,

0, ,

( , )

ci

i

k ik K

k i

a ak kk K

h g i I

h k K i I

v h a Aδ

∈

∈

= ∈

≥ ∈ ∈

= ∈

∑

∑


•Such models may be solved by a variety of algorithms. Often the sufficientconditions for convergence are impossible to verify

• A common heuristic method used in practice is the Method of SuccessiveAverages

•The “Step Size Mechanism” may be related to the averaging of the link costs orthe averaging of link flows

where T(xk) is the computed procedure that is used to obtain the next iterate

Network EquilibriumModels:Variational Inequality Formulations

x x x x x T xk kk

k k k k

kk

k

+ + +

=

∞

= + − =

< < − = +∞∑

1 1 1

1

0 1 1

α

α α α

( ) ( )

( )

;

;


A Complex Model:Rigorous Formulation-Heuristic Solution Algorithm

• Santiago, Chile• Complex demand model• Road Network Equilibrium• Transit Network Equilibrium• Combined modes: road-transit; transit-transit

Santiago, Chile Strategic Planning Model

• Base network

• 409 centroids including 49 parking locations

• 1808 nodes, 11,331 directional links

• 1116 transit lines and 52468 line segments

• 11 modes, including 4 combined modes

(bus-metro, txc-metro, auto-metro and auto passenger-metro)

• The demand

• subdivided into 13 socio-economic classes

• 3 trip purposes ( work, study, other )

• driving license holders can access to 11 modes

• no license holders can access to 9 modes

Base Network of Santiago, Chile

Variational Inequality Formulation

Find ( h T* *, ) ∈Ω such that

pn ij g G m ggp

rpnm

r rr

pn ijpng

ijpng

ijpng

gp

m gijpnm

ijpng

ijpnm

ijpnm

p

C h T h h T T T

T T T T h T

∑ ∑ ∑ ∑ ∑

∑⊆ ∈

∈

− + − +

− ≥ ∀ ∈

( )

* * * * *

* * *

[ ( , )( ) ln ( ))

ln( / )( ) , ( , ) .

φβ

φ

1

0 Ω

Trip Ends and Conservation of FlowConstraints

jij

pngipn

g

T O i p n∑∑ = ∀, , , ( )α ipn

inij

pngjp

g

T D j p∑∑∑ = ∀, , ( )ξ ip

T T j p n gijpnm

ijpng

m g

− = ∀∈

∑ 0, , , , ( )Lijpng

φ gp ( ) , , , ,h T j p n gr

pnmijpnm

r R m

− = ∀∈∑ 0 ( )µ ij

pnm

h r p n mrpnm ≥ ∀0, , , , ( )γ r

pnm

T ij p n mijpnm > ∀0, , , ,

T ij p n gijpng > ∀0, , , ,


Network Equilibrium Models: car and transit

• Multi-class network equlibrium model• Multi-class transit network equlibrium model• Heuristic equilibration that resorts to averaging of

flows and travel impedances

Solution Procedure

Multiple Transit Assignments

Standard Transit Assignment

for buses

Trip Distribution and Mode Choice

MetroMSA Impedance

Equil. AutoAssignment

Congested time

Equilibrium Transit Assignment

for metro

BusImpedance

AutoImpedance

Park-and-Ride Model for auto-metro and bus-metro

Convergence?

end All Impedances

Start and Initialize

MSA Auto Volume

Transit Autoequivalent flow

Auto Skims


Santiago, Chile Strategic Planning Model

• The next slide shows the convergence of the MSA algorithmthat uses link flow averaging for the car network and traveltime averaging for the transit network.

• The convergence of both the car demand and link flows aregiven for two variants: uncongested transit assignment andequilibrium transit assignment.

Convergence of equilibration

Normalized Gap vs. Iterationscongested vs. non-congested metro assignment with metro capacity

00.5

11.5

22.5

33.5

44.5

3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Iteration

Nor

mal

ized

gap

(%

)

cong. Link vol. cong. demand non-cong. Link vol. non-cong. demand

Auto demand

Auto link volume

Metro Volume (non-congested vs. congestedversion)

Metro Volume Changes

Auto-metro volume (non-congested vs.congested version)

Congested

Non-congested

Metro Volume - metro 5A, non-congested version

Metro Volum2 - metro 5A, congested version


Some Complex Applications

• Los Angeles, California• Toll Highways Poznan, Poland• Toll Bridge, Montreal,


The SCAG Regional TransportationPlanning Model

• A complex and very large scale model

• Lack of rigorous formulation; networkequilibrium sub-models

• A multi-class multi-mode network equilibriummodel with asymmetric cost function is part ofthe model

• Heuristic solution algorithm based on an outeraveraging scheme


SCAG MODELFLOW CHART:

START

auto skims forPK

auto skims forOP

transit skims forPK

transit skims forOP

HBW Logsums for PK (mc) HBW Logsums for OP (mc)

trip distribution for PK (gravity) trip distribution for OP (gravity)

mode choice model for PK mode choice model for OP

demand computations (time–of-day model)

auto-truck assignments for AM auto-truck assignments for MD

is the number of outer loops satisfied?

auto-truck assignments for PM auto-truck assignments for NT

transit assignments for AM transit assignments for MD

END

successive average link volume for outerloop of AM

successive average link volume for outerloop of MD

Trip generation


Network Overview - Highway Networkby facility type


Network Overview - Highway Networkwith parking lots


Equilibration Algorithm Convergence Results

SCAG Model Convergence (AM peak)

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1 5 9 13 2 6 10 14 3 7 11 15 4 8 12 1 5 9 13 17 21 25

loops

link

rela

tive

diffe

renc

e

inner-loop outer-loop


Assignment ResultsAM peak volume


Assignment ResultsAM link speed


Assignment ResultsAM truck volume by class


Assignment ResultsAM HOV VMT Grid Map


Assignment ResultsHCM Level of Service


Scenario ComparisonVMT changes


Toll highway analysis – Poznan, Poland

It involves the following models:

-Multi-class equilibrium assignment with generalizedcosts

-Demand models for toll-no toll choice and future yeardemands

-Equilibration of the demand for toll highway andnetwork performance


The multi-class equilibrium model withtolls

( )0

minav

c c ca a a

a A c C a As x dx v tθ

∈ ∈ ∈

+∑ ∑∑∫subject to , ,

0, ,

( , , )

ci

ci

ck i

k K

ck i

ca ak k

k K

h g i I c C

h k K i I

v h a A c Cδ

∈

∈

= ∈ ∈

≥ ∈ ∈

= ∈ ∈

∑

∑

khav

The numerical solution of this model is well known;It is worthwhile to point out that the flows by class, are not unique,nor are the path flows , but the arc flows are unique.

cav


Base year – No Toll


Base year – Medium Cost Toll


Base year – High Toll


Year 2010 – High Cost Toll


Year 2020 – Medium Cost Toll


Year 2010 – Low Cost Toll


Toll Bridge Study – Montreal, Canada

• Study carried out by the Ministry of Transportationof Quebec

• The analysis relied heavily on the analysis of pathsgenerated by the assignment algorithm


The Montreal Region


The new proposed bridge


Zones around Bridge


Demand for Current and Future Year


The Current Bridge Flows


Bridge Flows with New facility


Toll Income with Various Toll Levels


National Models

• These are very large multi-modal multi-class models• The underlying demand models are rather complex

and the running times are very high• An example of a national model is the PLANET

model developed for British Rail


The PLANET Model – British Rail

Zones Rail Road


Dynamic Network Equilibrium Model

• Solved by a hybrid optimization-simulation model adiscretized version of a variational inequality formulation of adynamic network equilibrium model

• The theoretical properties of the model are difficult to establish

• Wardrop’s user equilibrium in a temporal framework is a basisfor the model


( ) ( ) 0( )

( )

u t if h ti ks tk u ti

= >≥

otherwise ( ) ( )minu t s ti kk Ki

=∈

( ) ( )( )i

k ik K t

h t g t∀ ∈

=∑

Dynamic equilibrium

( ) 0kh t ≥

( )

( )( )

1,

i

k

t t T

i i I

g t i

s t t k

= ∈

= ∈

=

=

assignment interval

OD pair,

demand for OD pair

travel time ( ) on path

variables

constraints

equilibrium conditions

( )

( ) ( )

0,

i

k

T

I

K i

h t t k

=

=

=

=

demand period

set of OD pairs

set of paths for

flow on path


Dynamic assignment model

Networkdefinition

Time-dependentOD matrices

TrafficControl Data

STOP

calculate path flows (t)

run traffic simulation

choose initial paths

determine path traveltimes (t)

convergence achieved?YES NO

modify path sets:add a new path or

remove an existingpath (each O-D pair

& each interval)


Traffic simulation model

simplified model of vehicle interactions allows for anefficient event-based simulation

• car following• lane changing• gap acceptance

sophisticated lane selection heuristics• local lane selection rules• stochastic look-ahead strategy


Q

0

400

800

1200

1600

2000

0 90 K

flow

(veh

/hr/l

ane)

1

6030

-W

120

1

density (veh/km/lane)

V

free-flow speed

1

1

V

QL V R

K L

W L R

=

=

=

=

+

fundamental diagram


An application in Stockholm


An application in Stockholm


An application in Montreal


An application in Auckland,NZ


Ending Remarks

• The equilibrium model of route choice is hereto stay for both static and dynamic models

• We have a lot to thank to the landmarkcontribution of 1956

Documents

Network Equilibrium Models: Varied and Ambitious · Network Equilibrium Models: Varied and Ambitious Michael Florian Center for Research on Transportation University of Montreal