Preferred citation style - MATSim

Preferred citation style 

MATSim development team (ed.) (2007) MATSIM-T: Aims, 

approach and implementation, IVT, ETH Zürich, Zürich. 

1

MATSIM-T: Aims, approach and implementation 

August 2007

Overview 

• Structure and team 

• Task and solution methods 

• MATSIM aims 

• System architecture 

• Examples 

• Switzerland 

• Berlin/Brandenburg 

• Outlook and next steps 

• Progress on shortest path calculations 

• Traffic flow model 

• Improving the convergence 

3

Structure 

Software: 

• Open-source project under GNU public licence 

Coordination: 

Data: 

• Kai Nagel, TU Berlin 

• Public sources, where available 

• Private sources, when needed or as occasion arises 

4

Current team 

Strategy: 


• Kay Axhausen, ETH Zürich 

• Fabrice Marchal, LET, Lyon 

Coordination of the implementation and project management: 

• Michael Balmer, ETH Zürich 

• Marcel Rieser, TU Berlin 

5

Current team: Implementation (1/2) 

• Michael Balmer, ETH 

• David Charypar, ETH 

• Francesco Ciari, ETH 

• Dominik Grether, TU Berlin 

• Jeremy Hackney, ETH 

• Andreas Horni, ETH 

• Gregor Lämmel, TU Berlin 

• Nicolas Lefebvre, ETH 

• Michael Löchl, ETH 

6

Current team: Implementation (2/2) 

• Fabrice Marchal, LET 

• Konrad Meister, ETH 


• Marcel Rieser, TU Berlin 

• Nadine Schüssler, ETH 

• David Strippgen, TU Berlin 

7

Current funding sources 

• Basic research support for the chairs 

• (competitive) ETH research fund 

• Swiss National Fund 

• German Research Society 

• EU Framework funding 

• VW Foundation 

• Swiss Commission for Technology and Information (KTI) 

(datapuls, Lucerne) 

8

Task and solution methods 

9

What we would like to do ? 

Personal 

worlds of 

others 

Social captial: 

stock of joint 

abilities, 

shared 

histories and 

commitments 

Personal 

world 

Biography 

Projects Learning 

Household locations 

Social network geography 

Mobility tools 

10

What would we like to do ? Personal daily dynamics 

Physiological needs 

Commitments 

Desires 

Pending activities 

Scheduling 

Activity repertoire (t) ................ Activity repertoire (t+1) 

Activity 

calendar (t) 

Activity 

schedule (t) 

Rescheduling, 

Execution 

Unexecuted 

activities 

Networks, Opportunities 

Mental map (t) ................ Mental map (t+1) 

Updates, 

Innovations 

11

Understanding scheduling 

• Budget constraints 

• Capability constraints 

• Generalised costs of the schedule 

• Generalised cost of travel 

• Generalised cost of activity participation 

• Risk and comfort-adjusted weighted sums of time, 

expenditure and social content 

12

Degrees of freedom of activity scheduling 

• Number and type of activities 

• Sequence of activities 

• Start and duration of activity 

• Composition of the group undertaking the activity 

• Location of the activity 

• Connection between sequential locations 

• Location of access and egress from the mean of 

transport 

• Vehicle/means of transport 

• Route/service 

• Group travelling together 

13

Understanding supply 

Slot: A path in the time-space environment, which allows moving 

or activity performance 

• Regulated slots (e.g. table in a restaurant, reserved seat in a 

theatre, gate position of a plane, green light at a junction) 

• Emergent slots (e.g. trajectory of a car on a motorway, 

players in a pub-soccer tournament) 

Waiting time ~ Reserve capacity = Capacity – Demand for slots 

14

What do we (generally) do ? 

“Scenario” 

Competition for 

slots on networks 

and in facilities 

k(t,r,j) i,n 

q i _ (t,r,j) i,n 

Mental map 

Activity 

scheduling 

Population 

_i,t, r,j,k 

Demand q are the i th movements of person p from the current location 

at time t on route (connection) r to location j. The resulting 

generalised costs k are used to adjust the schedules and to change 

the capacities C and prices P of facilities f 

15

Classification criteria 

• Steady state (equilibrium) ? 

• Aggregate demands ? 

• Complete and perfect knowledge ? 

• Optimised schedules ? 

• Degrees of freedom and detail of scheduling 

• Modelling of capacity restrictions (movement, activities) ? 

16

MATSIM-T aims (1): Steady-state version 

• Steady state within 12 hours on a small multi-CPU machine 

• 7.5 mio agents, parcels, navigation networks 

• Shared time-of-day dependent generalised costs of travel 

and activity participation 

• Optimised schedules 

• Continuous time resolution; space: parcels; social networks 

• Queuing for slots for movement and activities 

17

MATSIM-T aims (2): Path-dependent version 

• Path-dependent development; precise estimates within 12h 

on a small multi-CPU machine 

• Large scale scenario 

• Agent-specific, learned time-of-day dependent generalised 

cost of travel and activity participation 

• Optimised schedules at multiple decision points 

• Continuous time resolution; space: parcels; social networks 

• Queuing for slots for movement and activities 

18

System architecture 

• XML standards 

• Data handling tools (aggregation/disaggregation) 

• Data base 

• Iteration handling and control of convergence 

• Models 

19

Architecture: JAVA - packages 

20

Architecture: Data flow 

JAVA C++ 

21

Current task allocation: Initial demand generation 








transport 




22

Current task allocation: activity scheduling 








transport 




23

Current ability: competition for slots 

Shortest paths: 

• A* dynamic algorythm 

Movement: 

• Queue-based simulation of car traffic 

• No representation of cycling, walking, public transport 

vehicles 

Activities 

• No competition for facilities yet 

24

Result of each iteration: Plan 

 

 

 

 

1900 1899 1897 

 

 

 

1899 1848 1925 1924 1923 1922 1068 

 

 

 

1067 1136 1137 1921 1922 1923 1925 1848 1899 

 

 

 

 

25

Scheduling and its utility function 

26

Utility function: Individual schedules 

U 

plan 

= 

n 

! 

i= 

1 

U 

act, 

i 

+ 

n 

! 

i= 

2 

,,.,actidurilateariUUU=+ 

U 

trav, 

i" 

1, 

i 

27

Utility function: Travel 

U 

trav, 

i&1, 

i 

= 

# 'trav 

% ttrav, 

i&1, 

i if ttrav, 

i&1, 

i $ 

" 

! 0 

else 

" 

trav 

= ! 12 

0 

Euro / h 

28

29 

Utility function: Late arrival 

( ) 

! 

" 

# $ 

% 

& 

= 

else 

t 

t 

if 

t 

t 

U 

i 

ar 

latest 

i 

start 

i 

ar 

latest 

i 

start 

ar 

late 

i 

ar 

late 

0 

, 

. 

, 

, 

. 

, 

. 

, 

. 

' 

h 

Euro 

ar 

late 

/ 

18 

. 

! 

= 

"

30 

Utility function: Activity performance 

( ) 

( ) 

! 

" 

! 

# 

$ 

> 

% 

< 

& 

& 

% 

% 

= 

0 

0 

0 

ln 

, 

, 

0 

, 

, 

, 

0 

, 

0 

, 

* 

, 

' 

' 

( 

else 

t 

t 

t 

if 

t 

t 

if 

t 

t 

t 

U 

i 

dur 

i 

i 

dur 

i 

dur 

i 

i 

i 

dur 

dur 

i 

dur 

h 

Euro 

h 

Euro 

dur 

/ 

6 

/ 

10 

= 

= 

! 

"

Scheduling: Current planomat(s) 

Version 1: 

• GA optimiser of durations and starting times 

• Retains time-of-day profile of generalised costs 

Version 2: 

• CMA-ES (Covariance Matrix Adaptation Evolutionary 

strategy) of durations and starting times 

• Retains time-of-day profile of generalised costs 

31

Examples (scenarios): 

• CH: Switzerland and subsets 

• Base tests for Kanton Zürich 

• B/B: Berlin/Brandenburg 

• Tolling case studies 

32

CH: Facilities for 140!000 hectares 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

33

CH: Comparisons at the aggregate levels 

34

CH: Network 

Simplification: 

Results: 

• Remove unconnected links and subnetworks 

• Remove nodes between unchanging link types 

• Cut off dead ends 

• Navteq (882!120 links) (25% loss of links/nodes) 

• Teleatlas (1!288!757 links) (currently not in use) 

35

CH: Population 

Alternatives: 

• Artificial sample generation from Census marginals 

• Census 

• Private census with additional imputations (datapuls, 

Lucerne) plus household “formation” 

36

CH: Mobility tools 

Approach: 

Data: 

• MNL of mobility tool packages (Beige) 

• Socio-demographics 

• Location types 

• Travel times to main centre by road and public transport 

• National travel survey (MZ 2000) 

37

CH: Structure of the chains, activity durations 

Data 

• MZ 2000 

• MZ 2005 

Approach: 

• Selection via conditional probability distribution from Chains 

* Person types frequencies 

• ca. 50 activity chains for MZ 2000 (>95% of all cases) 

• Specified with activity durations given by MZ 2000 

! xxxx different activity chains (weighted) 

38

CH: Activity location choice 

Home: 

• Random location in hectare 

Work/school 

• Random allocation from census commuter matrices 

• Disaggregation to facilities inside municipality 

Other locations 

• “Neighbourhood search” from given home/work/school 

locations (a variation of Gravity Model) 

39

CH: Mode choice 

Approach: 

Fixed mode choice at the tour level as a function of 

• Driving licence 

• Mobility tool ownership 

• Distance 

• Age * season ticket 

• Current set of chains include no subtours 

Data: 

• MZ 2005 

40

R: Hectare; M: Municipality 

CH: Kt. Zürich – Basic test 

41

CH: Kt. Zürich 

42

CH: Kanton Zürich 

43

CH: Kanton Zürich (planomat - Iteration 0) 

44

CH: Kanton Zürich (planomat - Iteration 400) 

45

B/B: Network 

• Source: “Senatsverwaltung für Stadtentwicklung Berlin” 

• 11596 nodes 

• 27770 links 

46

B/B: Population and demand 

• Generated from “Kutter-Nachfragemodell für Berlin”, 

• only commuters (work, education) in Berlin/Brandenburg, 

• only car-trips simulated, 

• 10% sample: 160!171 agents, each with 2 trips 

• Facilities = traffic analysis zones 

47

B/B: Tolling in Berlin/Brandenburg 

• Scenario: As described above 

• Aims: Impacts of distance and area tolls 

48

B/B: Tolling in Berlin/Brandenburg 

49

B/B: Trip length within toll area with distance tolls 

50

B/B: total trip lengths with distance tolls 

51

Outlook 

52

Current tasks: Functionality 

• Improving the realism of the scenario (e.g. opening hours, 

parameter distributions) 

• Validation of the current scenarios 

• Switzerland scenario in 12h to steady state 

• Functional expansion the planomat (mode choice, 

destination choice) 

• Parameter estimation for planomat 

• Interface to UrbanSim 

53

Future tasks: Functionality 

• Integration of social network data structures 

• Explicit social network choices 

• Addition of supply agents (car sharing, demand responsive 

transport, retail location, parking pricing, road pricing) (Traffic 

control) 

54

Current tasks: Usability (Shareability) 

• Analysis tools 

• Anonymous test data sets 

• Improved documentation 

• Web and sourceforge maintenance 

• User support, including visits to ETH/TU Berlin 

55

Acknowledgements (1/3) 

System architecture: Michael Balmer 

System integration: Michael Balmer, Marcel Rieser 

System management: Michael Balmer, Marcel Rieser 

Facilities: Konrad Meister 

Mobility tools: Francesco Ciari 

Tour mode choice: Francesco Ciari 

Counter infrastructure: Andreas Horni 

KML – interfaces: Konrad Meister, Dominik Grether 

56


Router: Nicolas Lefebvre 

Traffic flow model: David Charypar 

Planomat: Konrad Meister, David Charypar 

Supply agents: Michael Löchl, Francesco Ciari 

Social networks: Jeremy Hackney 

57


Past contributors: 

• Michael Bernard, ETH 

• Sigrun Beige, ETH 

• Ulrike Beuck, TU-Berlin 

• Dr. Nurhan Cetin (ETH) 

• Phillip Fröhlich, Modus, Zürich (ETH) 

• Dr. Christian Gloor (ETH) 

• Dr. Bryan Raney, HP America (ETH) 

• Marc Schmitt (ETH) 

• Arnd Vogel, PTV (TU Berlin) 

58

More information 

www.matsim.org 

www.vsp.tu-berlin.de 

www.ivt.ethz.ch/vpl/publications/reports 

59

Progress on shortest – path calculation 

Possibilities: 

" Bounding boxes: Find out whether certain nodes can at all be on 

a shortest path and if not, do not consider them. 

" Multi-level approach: Add shortcuts to the network where 

possible to bypass several edges at a time when routing. 

" Bi-directed search: Start routing at the end and at the start node 

at the same time. 

" Goal-directed search: Change the way the nodes are ranked such 

that nodes which are less likely to be on the shortest path are also 

less likely to be visited. The most popular algorithm that uses this 

technique is A*. 

60

Speed-up 

Basic algorithm: A* 

Step 1: Improve data structure for list of candidate nodes (7% 

reduction) 

Step 2: Improve handling of the “visited flag” (~ length of route) 

Step 3: Detect dead ends (1/4 of nodes in the NavTeq network) 

(50% reduction) 

Step 4: Use euclidian-distance to destination to rank candidates 

(50-80% reduction) 

Step 5: Use intermediate landmarks (80-90% reduction) 

61

Landmark selection 

• Divide network into sectors 

containing an equal number 

of nodes 

• For each sector, choose 

a node that is farthest 

away from the center 

• Check that the 

landmarks are not 

too close to each 

other. If so, narrow 

one sector and choose 

a new landmark within the sector 

62

Speed up of shortest-path calculations 

Free flow conditions: 97-99% reductions 

Loaded conditions: 95% reductions 

on navigation networks 

63

Traffic flow model 

“Scenario” 




k(t,r,j) i,n 

q i _ (t,r,j) i,n 

Mental map 

Activity 

scheduling 

Population 

_i,t, r,j,k 

64

Approach 

Resolution 

physical 

(VISSIM) 

CA 

(TRANSIMS) 

Q 

(Cetin) 

Q event 

(MATSIM) 

meso 

(METROPOLIS) 

parallel 

Q event 

(MATSIM) 

macro 

(VISUM) 

Speed 

65

Parallel q-event driven simulation with gaps 

• Approach 

• Fundamental diagram (on a ring motorway) 

• Domain decomposition 

• Test 

66

Q-event: Approach without gaps 

t 

Link 1 Link 2 

v free

Q-event: Approach with gaps 

t 

Link 1 Link 2 

v gap

Q-event: Implementation details 

• Squeezing to avoid grid-lock 

• Inflow capacity = 110% of outflow capacity (1800 veh/h* 

lanes) 

• Vehicles are served in order of arrival at the junctions 

• C++ with binary data interface to MATSim-T 

69

Q-event: Fundamental diagram 

70

Q-event: Test setup 

• Road network of the federal states of Germany Berlin and 

Brandenburg 

• 11.6k nodes, 27.7k links 

• 7.05M person days 

• Average number of trips per agent: 2.02 

• Average length of a trip: 17.5 links 

• Total: 249M road segments to be traveled 

• 77 min on a single dual-core CPU (1.6GH; 256 GB RAM)

Q-event: Integrated domain decomposition 

72

Q-event: Parallelisation

Improving the convergence 

74

Initial approach 

Method: 

• Plan everybody at Iteration 1 

• Replan for a fixed share 

Convergence: 

• On mean performance 

75

Improved approach 

Method: 

• Plan for everyone at Iteration 1 

• Replan for a predetermined, but decreasing share over the 

number of iterations 

Measurement: 

• Aggregates 

76

Impact of improved approach 

77

Support graphs and further ideas 

78

What should we do, when we do what we do ? 




k(t,r,j) i,n 

q i _ (t,r,j) i,n 

Activity 

scheduling 

_ i,t, r,j,k 

Observed 

schedules and 

generalised costs 

Parameter 

calibration 

79

What would we like to do ? 

Locating, sizing 

and pricing of 

slots 

k(t,r,j) i,n 

C f (t); P f (t) 




k(t,r,j) i,n 

q _ (t,r,j) i,n 

Mental map 

Activity scheduling 

(private, 

commercial) 

80

What would we like to do ? Personal long-term dynamics 

Definition of „Self“ 

Desires 

Pending projects 

Planning, 

Negotiation 

(Life) goals (t) ................ (Life) goals (t+1) 

Projects (t) 

[committments] 

Project 

sequence (t) 

Replanning, 

Execution 

Unexecuted 

projects 

Markets and networks 

Updates, 

Innovations, 

Reflection 

Personal world (t) ................ Personal world (t+1) 

81

Architecture: Initial data preparation 

82

Architecture: Learning, search for consistency 

83

CH: Car availability (Census) 

84

CH: Mode choice – Observed share public transport 

85

Scheduling: Once and future Planomat 

• Number and type of activities out of an agenda 



• Composition of the household group undertaking the activity 




transport 




86

CH: Facilities – Data sources 

Three tables of Enterprise Cernsus: 

NOGA Description Firms FT-equivalents 

Code 

10-45 (sector 2) 6 200 

19 Leather and shoe production 2 0-9 

20 Chemical industry 3 50-249 

19.10 Leather production yes 

19.20 Leather goods no 

19.30 Shoe production no 

87

CH: Facilities – current allocation 

• Read census of employment for each hectare 

• Generate the required number of facilities 

• Add activity work 

• Set number of work place to minimum of class 

• Distribute remainder proportionally to class size 

• Add activity of use 

• Add standard opening hours 

• Randomly distribute on the hectare and attach to nearest link 

88

CH: Facilities - results 

89


90


91

Preferred citation style - MATSim

Create successful ePaper yourself

Delete template?

Save as template?