Large-scale agent-based transport simulations in France

Sebastian Hörl

17 October 2025

Presentation at CIRED

The street in 1900

http://www.loc.gov/pictures/item/2016800172/

The street today

https://commons.wikimedia.org/wiki/File:Atlanta_75.85.jpg

The street of tomorrow?

Julius Bär / Farner

Macroscopic transport modeling

Classic four-step models

  • Travel demand generated in (large) zones
  • Focus on large flows between these zones
  • For the morning or evening commute
  • For limited user groups
     
  • Question: Where to add capacity?

Macroscopic transport modeling

Classic four-step models

  • Travel demand generated in (large) zones
  • Focus on large flows between these zones
  • For the morning or evening commute
  • For limited user groups
     
  • Question: Where to add capacity?

Macroscopic transport modeling

Classic four-step models

  • Travel demand generated in (large) zones
  • Focus on large flows between these zones
  • For the morning or evening commute
  • For limited user groups
     
  • Question: Where to add capacity?

Agent-based transport modeling

0:00 - 8:00

08:30 - 17:00

17:30 - 0:00

0:00 - 9:00

10:00 - 17:30

17:45 - 21:00

22:00 - 0:00

  • Individual travellers with daily activities
  • Moving from one activity to another
  • Simulation of the entire day
  • Highly detailed interaction between travellers and services
     
  • Multitude of (design) questions can be answered

Agent-based transport modeling

How to set up agent-based transport simulations?

* with reproducible results

* in a replicable way

Transport modeling chain

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Transport modeling chain

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Synthetic populations: Introduction

Definition

  • Representation digital version of the real population of a territory
  • Persons (single-level) or households with persons (two-level) population
  • Households and persons with individual attributes
  • Persons with individual activity chains

0:00 - 8:00

08:30 - 17:00

0:00 - 9:00

10:00 - 17:30

17:45 - 21:00

22:00 - 0:00

17:30 - 0:00

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

Data

Goals

  • Generate individual households and persons
  • Choose a distinct place of residence

French population census

Household ID Person ID Zone Age Sex ... Weight
512 1 75013 35 f ... 3.2
512 2 75013 32 m ... 3.2
516 1 75019 42 m ... 4.1
... ... ... ... ... ...

Upsampling of persons using Truncate-Replicate-Sample (TRS)

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

Data

Goals

  • Generate individual households and persons
  • Choose a distinct place of residence

Sampling by number of housing units per building

French bulding database

French address database

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

Data

Goal

  • Choose work places and education locations

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

French work
commuting matrix

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

Data

Goal

  • Choose work places and education locations

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

French work
commuting matrix

National enterprise 
database

with facilities by number of employees

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

Data

Goal

  • Choose work places and education locations

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

French education
commuting matrix

Permanent facility
database

with education facilities and attendants

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

Data

Goal

  • Generate activity sequences (type, start and end time) for each person

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

Statistical Matching

National Household Travel Survey 2008

(Local Household Travel Surveys)

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

Data

Goals

  • Choose locations of secondary (shopping, leisure, ...) activities

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

Hörl, S., Axhausen, K.W., 2021. Relaxation–discretization algorithm for spatially constrained secondary location assignment. Transportmetrica A: Transport Science 1–20.

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

Data

Output

  • Three main tables: households, persons, activities
  • Supplementary tables: commutes, trips, ...

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

household_id income number_of_cars ...
1024 85,000 2 ...
household_id person_id age sex employed ...
1024 1 34 f true ...
1024 2 36 m true ...
household_id person_id activity_id start_time end_time type location ...
1024 1 1 00:00 08:00 home (x, y) ...
1024 1 2 09:00 18:00 work (x, y) ...
1024 1 3 19:00 24:00 home (x, y) ...

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

Data

Output

  • Three main tables: households, persons, activities
  • Supplementary tables: commutes, trips, ...

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

Place of residence

Commuting trips

Hourly work activities

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

Data

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

Validation

  • Comparison with census data, HTS data, ...

Synthetic populations: Pipeline

Pipeline

  1. Households and persons
  2. Primary activity locations
  3. Activity chains
  4. Secondary activity locations

Data

Validation

  • Comparison with census data, HTS data, ...

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

Synthetic populations: Pipeline

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

Open data

Open source

+

=

Replicable research in agent-based transport simulation

eqasim-org/ile-de-france

Synthetic populations: Pipeline

RP

BAN

BD-TOPO

MOBPRO

MOBETUD

SIRENE

BPE

ENTD

Open data

Open source

+

=

Lille

Paris

Strasbourg

Lyon

Toulouse

Bordeaux

Nantes

Rennes

Contributors

Users

Synthetic populations: Community

eqasim workshop in November 2025

Synthetic populations: Adaptations

Screenshot Sao Paolo

Copy & paste of the code base

Difficulty of maintenance

São Paulo

Almost same open data available as in France

California

Substantial modifiations required

Switzerland

Not based on open data
(for now)

Paper published in
Regional Studies, Regional Science (2020)

Paper presented at the Annual Meeting of the Transportation Research Board (2021)

Work in progress at ETH Zurich

Synthetic populations: Adaptations

Cairo: Extreme case, very few data available and not in the right format

Idea: Use data to generate "fake" input to the French pipeline and reuse the code!

Gall, T., Vallet, F., Reyes Madrigal, L.M., Hörl, S., Abdin, A., Chouaki, T., Puchinger, J., 2023. Sustainable Urban Mobility Futures, Sustainable Urban Futures. Springer Nature Switzerland, Cham.

Synthetic populations: Adaptations

Cairo: Extreme case, very few data available and not in the right format

Idea: Use data to generate "fake" input to the French pipeline and reuse the code!

Bavaria: Set up a robust and replicable pipeline with data replacement

Hörl, S., Burianne, A., Natterer, E., Engelhardt, R., Müller, J. (2025) Towards a replicable synthetic population and agent-based transport model for Bavaria, paper presented at the 23rd International Conference on Practical applications of Agents and Multi-Agent Systems (PAAMS 2025), June 2025, Lille, France.

As part of the national project MINGA

Synthetic populations: Adaptations

Cairo: Extreme case, very few data available and not in the right format

Idea: Use data to generate "fake" input to the French pipeline and reuse the code!

Bavaria: Set up a robust and replicable pipeline with data replacement

As part of the national project MINGA

Hörl, S., Burianne, A., Natterer, E., Engelhardt, R., Müller, J. (2025) Towards a replicable synthetic population and agent-based transport model for Bavaria, paper presented at the 23rd International Conference on Practical applications of Agents and Multi-Agent Systems (PAAMS 2025), June 2025, Lille, France.

eqasim-org/bavaria

Synthetic populations: Outlook

Improvements of models along the synthesis chain

  • In each step, models can be improved by more intelligent approaches from machine learning, deep learning, ...
  • How can we benchmark them against each other?

Uncertainty analysis

  • What impact does the replacement of one model have on the final output?
  • Which steps cause the highest degree of uncertainty? Which ones should be updated with priority?

Context-sensitive population synthesis

  • How can we make algorithms along the chain context-sensitive (infrastructure, access to work places, nature, ...)

Primary locations

Activity chains

Secondary locations

Persons

Complexity of synthetic populations

  • Weekly or longer activity sequences
  • Activity chains synchronized across household members

Synthetic populations: Scaling up

Use cases

  • Scaling to new use cases (see Cairo, Munich)
  • Standardize methods and computational tools

Application domains

  • Today mostly use in transportation
  • Large interest in the modeling of energy systems
  • Further domains: disaster management, climate change adaption, ...

Transport modeling chain

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Transport modeling chain

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Agent-based simulation: Introduction

GTFS

OpenStreetMap

Synthetic demand

+

Driving car

Metro / Train

Work activity starts

Agent-based simulation: Introduction

Synthetic demand

Agent-based simulation: Introduction

Mobility simulation

Synthetic demand

Daily mobility plans

Agent-based simulation: Introduction

Decision-making

Mobility simulation

Synthetic demand

Experienced travel times, crowding, ...

Daily mobility plans

Agent-based simulation: Introduction

Decision-making

Mobility simulation

Synthetic demand

Experienced travel times, crowding, ...

Daily mobility plans

Agent-based simulation: Introduction

Decision-making

Mobility simulation

Synthetic demand

Experienced travel times, crowding, ...

Daily mobility plans

Update

Agent-based simulation: Introduction

Decision-making

Mobility simulation

Synthetic demand

  • Maintained by TU Berlin, ETH Zurich, (IRT SystemX)
  • 50+ research users world-wide, SBB, Volkswagen, ...
  • Contributor since ~2016

Mode shares

Traffic patterns

Emissions

Noise

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic

Agent-based simulation: eqasim-java

Multi-modal approach

Mobility simulation

Queue-based network simulation

  • Car drivers (and passengers)
  • Public transport
  • Bicycle
  • Walking

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic

1) Simulation using queue model

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic

1) Simulation using queue model

2) Tracking of emerging travel times for each link k in time bin τ

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic

1) Simulation using queue model

2) Tracking of emerging travel times for each link k in time bin τ

3) Use tracked travel times for routing in next decision-making step

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Traffic

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Traffic

n_{k,\tau}

1) Track flow in time bin τ for each link k

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Traffic

n_{k,\tau}
t_{k,\tau} = t_{k,0} \left( 1 + a \left( \frac{n_{k,\tau}}{c_{k}} \right)^b \right)

1) Track flow in time bin τ for each link k

2) Calculate travel time according to VDF

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Traffic

n_{k,\tau}
t_{k,\tau} = t_{k,0} \left( 1 + a \left( \frac{n_{k,\tau}}{c_{k}} \right)^b \right)

1) Track flow in time bin τ for each link k

2) Calculate travel time according to VDF

3) Use updated travel times during routing and decision-making

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Traffic

n_{k,\tau}
t_{k,\tau} = t_{k,0} \left( 1 + a \left( \frac{n_{k,\tau}}{c_{k}} \right)^b \right)

1) Track flow in time bin τ for each link k

2) Calculate travel time according to VDF

3) Use updated travel times during routing and decision-making

4) Impose updated travel times in next iteration's traffic

Chouaki, T., & Hörl, S. (2025, January). A method for efficiently assessing the impact of local mobility services in large-scale agent-based simulations. 104th Annual Meeting of the Transportation Research Board (TRB 2025). 104th Annual Meeting of the Transportation Research Board (TRB 2025), Washington D.C., USA.

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Co-evolutionary decision-making algorithm

Traffic

Decision-making

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Co-evolutionary decision-making algorithm

Traffic

Decision-making

Score

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Co-evolutionary decision-making algorithm

Traffic

Decision-making

Score

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Co-evolutionary decision-making algorithm

Discrete choice models for behavioral simulation

Traffic

Decision-making

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Co-evolutionary decision-making algorithm

Discrete choice models for behavioral simulation

Traffic

Decision-making

P[k] = \frac{\exp\left(u_k \right)}{\sum_{k'}\exp\left( u_{k'} \right)}
u_\text{car} = \beta_\text{asc,car}
+ \beta_\text{travelTime,car} \cdot x_\text{travelTime,car}
+ \beta_\text{cost} \cdot x_\text{cost,car}
u_\text{pt} = \beta_\text{asc,pt}
+ \beta_\text{inVehicleTime,pt} \cdot x_\text{inVehicleTime,pt}
+ \beta_\text{pt} \cdot x_\text{cost,pt}
+ \beta_\text{waitTime,pt} \cdot x_\text{waitTime,pt}
+ \beta_\text{transfers,pt} \cdot n_\text{transfers,pt}

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Co-evolutionary decision-making algorithm

Discrete choice models for behavioral simulation

Traffic

Decision-making

P[k] = \frac{\exp\left(u_k \right)}{\sum_{k'}\exp\left( u_{k'} \right)}
u_\text{car} = \beta_\text{asc,car}
+ \beta_\text{travelTime,car} \cdot x_\text{travelTime,car}
+ \beta_\text{cost} \cdot x_\text{cost,car}
u_\text{pt} = \beta_\text{asc,pt}
+ \beta_\text{inVehicleTime,pt} \cdot x_\text{inVehicleTime,pt}
+ \beta_\text{pt} \cdot x_\text{cost,pt}
+ \beta_\text{waitTime,pt} \cdot x_\text{waitTime,pt}
+ \beta_\text{transfers,pt} \cdot n_\text{transfers,pt}

Estimate choice variables (routing)

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Co-evolutionary decision-making algorithm

Discrete choice models for behavioral simulation

Traffic

Decision-making

P[k] = \frac{\exp\left(u_k \right)}{\sum_{k'}\exp\left( u_{k'} \right)}
u_\text{car} = \beta_\text{asc,car}
+ \beta_\text{travelTime,car} \cdot x_\text{travelTime,car}
+ \beta_\text{cost} \cdot x_\text{cost,car}
u_\text{pt} = \beta_\text{asc,pt}
+ \beta_\text{inVehicleTime,pt} \cdot x_\text{inVehicleTime,pt}
+ \beta_\text{pt} \cdot x_\text{cost,pt}
+ \beta_\text{waitTime,pt} \cdot x_\text{waitTime,pt}
+ \beta_\text{transfers,pt} \cdot n_\text{transfers,pt}

Sampling

Hörl, S., & Balac, M. (2021). Introducing the eqasim pipeline: From raw data to agent-based transport simulation. Procedia Computer Science, 184, 712–719.

Hörl, S., Balac, M., & Axhausen, K., W. (2019, January). Pairing discrete mode choice models and agent-based transport simulation with MATSim. 98th Annual Meeting of the Transportation Research Board (TRB 2019). 98th Annual Meeting of the Transportation Research Board (TRB 2019), Washington D.C., USA.

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Co-evolutionary decision-making algorithm

Discrete choice models for behavioral simulation

Traffic

Decision-making

P[k] = \frac{\exp\left(u_k \right)}{\sum_{k'}\exp\left( u_{k'} \right)}
u_\text{car} = \beta_\text{asc,car}
+ \beta_\text{travelTime,car} \cdot x_\text{travelTime,car}
+ \beta_\text{cost} \cdot x_\text{cost,car}
u_\text{pt} = \beta_\text{asc,pt}
+ \beta_\text{inVehicleTime,pt} \cdot x_\text{inVehicleTime,pt}
+ \beta_\text{pt} \cdot x_\text{cost,pt}
+ \beta_\text{waitTime,pt} \cdot x_\text{waitTime,pt}
+ \beta_\text{transfers,pt} \cdot n_\text{transfers,pt}

Sampling

* tour-based

Hörl, S., & Balac, M. (2021). Introducing the eqasim pipeline: From raw data to agent-based transport simulation. Procedia Computer Science, 184, 712–719.

Hörl, S., Balac, M., & Axhausen, K., W. (2019, January). Pairing discrete mode choice models and agent-based transport simulation with MATSim. 98th Annual Meeting of the Transportation Research Board (TRB 2019). 98th Annual Meeting of the Transportation Research Board (TRB 2019), Washington D.C., USA.

Agent-based simulation: Discrete choice integration

Discrete choice-based decision-making

  • Estimation (computational effort; correctness)
  • Only mode choice in the current form
     
  • Realistic decisions with high probability
  • "Use the model as is" is advantage in calibration

Mobility simulation

Mutation / Selection

Mobility simulation

Mode choice

Estimation

Simulation stabilized?

Simulation stabilized?

Estimation correct?

Scoring-based decision-making

  • Very flexible (departure time, ...)
     
  • "Stupid" decisions necessary
     
  • Tedious calibration of parameters

Agent-based simulation: eqasim-java

eqasim-java: A streamlined set-up of MATSim for our standardized synthetic populations

eqasim

Detailed queue-based traffic simulation

Traffic based on volume-delay functions

Co-evolutionary decision-making algorithm

Discrete choice models for behavioral simulation

Standardized calibration procedure for France

Traffic

Decision-making

Agent-based simulation: Calibration

Capacity calibration

Household travel survey

Idea:

- Provide an open and accessible, well-calibrated agent-based simulation for Île-de-France

- Develop a standardized pipeline for the estimation of the discrete choice model and network calibration

Status:

- Under development in bits and pieces since ~2023

- Île-de-France simulation almost ready to be published!

EGT

Agent-based simulation: Calibration

Transit calibration

Capacity calibration

Household travel survey

Input: All trips of the survey that have been performed using public transport, with origin, destination and departure time.

EGT

Approach: Calibrate routing parameters of public transport routing algorithm (RAPTOR) to fit

  (1) Distribution of legs by public transport mode

  (2) Distribution of transfer counts

Adaptation of parameters

Distribution modes

Distribution transfers

Agent-based simulation: Calibration

Transit calibration

Freeflow calibration

Capacity calibration

Household travel survey

EGT

Input:
- A representative selection of car trips from the survey including origin, destination (and weight)

- Road network with speed limit travel times

Approach:

  (1) Route trips using API (TomTom) without congestion (3am)

  (2) Calibrate "crossing penalties" at nodes by road type

Agent-based simulation: Calibration

Transit calibration

Freeflow calibration

Capacity calibration

Household travel survey

EGT

Input:
- A representative selection of car trips from the survey including origin, destination (and weight)

- Road network with speed limit travel times

Approach:

  (1) Route trips using API (TomTom) without congestion (3am)

  (2) Calibrate "crossing penalties" at nodes by road type

Agent-based simulation: Calibration

Transit calibration

Freeflow calibration

Congestion model

Capacity calibration

Household travel survey

EGT

Input: A representative selection of car trips from the survey including origin, destination, departure time, and calculated free-flow travel times

Approach:

  (1) Route trips using API (TomTom) with congestion (at trip departure time)

  (2) Calibrate hourly correction factors

t_{k,c} = t_{k,0} \cdot \prod_{p \in \text{pred(k)}} f_{h,p}
\log t_{k,c} = \log t_{k,0} + \sum \delta_{k,p} \cdot \log f_{h,p}

Linear regression

Agent-based simulation: Calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

EGT

Approach:

  (1) Route all trips of the survey for all mode alternatives
  (2) Estimate a multi-nomial logit model using Biogeme

  • Automatic construction of zonal dummies (if requested)
  • Modes: Car driver, car passenger, public transport, bicycle, walking
  • Currently R2 = 0.52
     
  • Value of time comparable to existing studies (Meunier & Quinet, ...)
  • Automatic calculation and reporting of elasticities

Agent-based simulation: Calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

EGT

Approach:

  (1) Route all trips of the survey for all mode alternatives
  (2) Estimate a multi-nomial logit model using Biogeme

Focus: Parking model

  • Parking costs obtained across Île-de-France from open web sources
     
  • Calculation of "parking pressure"
  • High model impact!

Registered vehicles inside IRIS

Road network inside IRIS

P =

* from INSEE data

Agent-based simulation: Calibration

Capacity calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

EGT

Approach:

  (1) Implement choice model in agent-based simulation

  (2) Verify point-to-point travel times between simulation and API

  (3) Adjust a (global) capacity factor across the network

Comparison

Simulation

Capacity factor

Point-to-point travel times

* no flow comparison!

Agent-based simulation: Calibration

Capacity calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

EGT

Validation:

  - Global mode share

  - Mode share by distance

  - CDF of travel times by mode
  - Others possible; some within limits (traffic counts, transit counts, ...)

 

* not the latest results

Agent-based simulation: Calibration

Capacity calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

EGT

Validation:

  - Global mode share

  - Mode share by distance

  - CDF of travel times by mode
  - Others possible; some within limits (traffic counts, transit counts, ...)

 

* not the latest results

Agent-based simulation: Calibration

Capacity calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

EGT

Validation:

  - Global mode share

  - Mode share by distance

  - CDF of travel times by mode
  - Others possible; some within limits (traffic counts, transit counts, ...)

 

* not the latest results

Agent-based simulation: Calibration

Capacity calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

EGT

Validation:

  - Global mode share

  - Mode share by distance

  - CDF of travel times by mode
  - Others possible; some within limits (traffic counts, transit counts, ...)

 

* not the latest results

Agent-based simulation: Calibration

Capacity calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

EGT

Baseline simulation for Île-de-France ready to publish!

 

Fully automatized and replicable processing pipeline using snakemake

Capacity calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

EDGT, EMC2, ...

Agent-based simulation: Outlook

Can this be generalized to other cases?

Most likely!

!

Capacity calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

EDGT, EMC2, ...

Spatialization

Agent-based simulation: Outlook

Can this be generalized to other cases?

Most likely!

Balac, M., Hörl, S., & Schmid, B. (2024). Discrete choice modeling with anonymized data. Transportation, 51(2), 351–370.

Capacity calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

Spatialization

Lille

Paris

Strasbourg

Lyon

Toulouse

Bordeaux

Nantes

Rennes

Compare individual model results based on same structure

Agent-based simulation: Outlook

EDGT, EMC2, ...

Agent-based simulation: Outlook

Capacity calibration

Transit calibration

Freeflow calibration

Congestion model

Model estimation

Capacity calibration

Household travel survey

Spatialization

Lille

Paris

Strasbourg

Lyon

Toulouse

Bordeaux

Nantes

Rennes

Estimate a joint model for France?

EDGT, EMC2, ...

Transport modeling chain

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Transport modeling chain

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Automated taxi

Pickup

Dropoff

Use cases: On-demand mobility

  • An operator centrally controls a fleet of vehicles
     
  • Each vehicle is represented as an agent that receives instructions in each time step
     
  • Customer agents sent requests to be transported
     
  • Objectives: maximize operator revenue, minimize empty distance, ...

Automated taxi

Pickup

Dropoff

Use cases: On-demand mobility

  • An operator centrally controls a fleet of vehicles
     
  • Each vehicle is represented as an agent that receives instructions in each time step
     
  • Customer agents sent requests to be transported
     
  • Objectives: maximize operator revenue, minimize empty distance, ...

R1

Use cases: On-demand mobility

  • Different dispatching strategies provide different outcomes in terms of empty distance, revenue, and wait times

amodeus-science/amodeus

AI Driving Olympics challenge at NeurIPS 2018

Use cases: On-demand mobility

Cost model

Discrete choice model

Mobility simulation

Estimation

Fare per trip and km

Wait time

Outcomes

Passenger distance, empty distance

  • The problem becomes even more interesting when customer agents have the choice to use the service or not (dynamic demand)

Use cases: On-demand mobility

  • Provides an understanding of a mobility service that doesn't exist today
     
  • Shows pathways for policy and regulation

Hörl, S., Becker, F., & Axhausen, K. W. (2021). Simulation of price, customer behaviour and system impact for a cost-covering automated taxi system in Zurich. Transportation Research Part C: Emerging Technologies, 123, 102974.

Use cases: On-demand mobility

Hörl, S., Balac, M., & Axhausen, K. W. (2019). Dynamic demand estimation for an AMoD system in Paris. IEEE Intelligent Vehicles Symposium (IV 2019), 260–266.

On-demand mobility: Intermodality

  • How to combine on-demand mobility systems with public transport?

     


  • How to take into account rejection rates in discrete choice models?





  • Various other publications ...

Chouaki, T., Hörl, S., Puchinger, J., 2023. Towards Reproducible Simulations of the Grand Paris Express and On-Demand Feeder Services, in: 102nd Annual Meeting of the Transportation Research Board (TRB 2023). Washington D.C, United States.

Chouaki, T., Hörl, S., Puchinger, J., 2023. Control-based integration of rejection rates into endogenous demand ride-pooling simulations, in: 8th International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS 2023). IEEE, Nice, France, pp. 1–6.

  • How to combine on-demand mobility systems with public transport?

     

     
  • How to take into account rejection rates in discrete choice models?




     
  • Various other publications ...

Chouaki, T., Hörl, S., Puchinger, J., 2023. Towards Reproducible Simulations of the Grand Paris Express and On-Demand Feeder Services, in: 102nd Annual Meeting of the Transportation Research Board (TRB 2023). Washington D.C, United States.

Chouaki, T., Hörl, S., Puchinger, J., 2023. Control-based integration of rejection rates into endogenous demand ride-pooling simulations, in: 8th International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS 2023). IEEE, Nice, France, pp. 1–6.

On-demand mobility: Intermodality

  • How to combine on-demand mobility systems with public transport?

     

     
  • How to take into account rejection rates in discrete choice models?




     
  • Various other publications ...

Chouaki, T., Hörl, S., Puchinger, J., 2023. Towards Reproducible Simulations of the Grand Paris Express and On-Demand Feeder Services, in: 102nd Annual Meeting of the Transportation Research Board (TRB 2023). Washington D.C, United States.

Chouaki, T., Hörl, S., Puchinger, J., 2023. Control-based integration of rejection rates into endogenous demand ride-pooling simulations, in: 8th International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS 2023). IEEE, Nice, France, pp. 1–6.

On-demand mobility: Intermodality

On-demand mobility: Algorithmic fairness

Do dispatching algorithms discriminate against certain user groups?

  • Standard algorithms aim at minimizing wait times, travel times and maximizing revenue
     
  • Do standard algorithms reject mobility-impaired person with longer interactions or larger groups more frequently than others?
     
  • Yes, they do!




     
  • Can we mitigate the problem?
     
  • Opens a whole new section of research in fleet management

Chouaki, T., Hörl, S., 2024. Comparative assessment of fairness in on-demand fleet management algorithms, in: The 12th Symposium of the European Association for Research in Transportation (hEART). Espoo, Finland.

Infrastructure: Charging behaviour

How do people choose between public, home and work chargers for their electric cars?
 

  • Very few data available (surveys and use)
     
  • Idea 
    - Assign electric vehicles to the population, then force them to charge (to avoid zero SoC)
    - What is their ideal charging configuration, given the provided infrastructure?
     
  • Collective charging strategy selection process (home, work, public) through maximization of scores
     
  • Negative scores for zero SoC, falling below a minimum SoC during the day or at the end, monetary costs, ...

Infrastructure: Charging behaviour

How do people choose between public, home and work chargers for their electric cars?
 

  • Very few data available (surveys and use)
     
  • Idea 
    - Assign electric vehicles to the population, then force them to charge (to avoid zero SoC)
    - What is their ideal charging configuration, given the provided infrastructure?
     
  • Collective charging strategy selection process (home, work, public) through maximization of scores
     
  • Negative scores for zero SoC, falling below a minimum SoC during the day or at the end, monetary costs, ...

Transport policy: Limited traffic zones

What is the impact of the Limited Traffic Zone in the center of Paris?
 

  • Rule: Non-residents that are not performing an activity in the center of Paris are not allowed to go through the center zone
     
  • We can analyze which persons (agents) are affected by that policy
     
  • We can measure the impact of the policy on the surrounding traffic
     
  • High level estimation of traffic and emission impact

Residents

Transit

Visitors

Transport policy: Limited traffic zones

What is the impact of the Limited Traffic Zone in the center of Paris?
 

  • Rule: Non-residents that are not performing an activity in the center of Paris are not allowed to go through the center zone
     
  • We can analyze which persons (agents) are affected by that policy
     
  • We can measure the impact of the policy on the surrounding traffic
     
  • High level estimation of traffic and emission impact

Transport policy: Limited traffic zones

What is the impact of the Limited Traffic Zone in the center of Paris?
 

  • Rule: Non-residents that are not performing an activity in the center of Paris are not allowed to go through the center zone
     
  • We can analyze which persons (agents) are affected by that policy
     
  • We can measure the impact of the policy on the surrounding traffic
     
  • High level estimation of traffic and emission impact

Overall flow related to the ZTL

Transport policy: Limited traffic zones

Transit flow related to the ZTL

What is the impact of the Limited Traffic Zone in the center of Paris?
 

  • Rule: Non-residents that are not performing an activity in the center of Paris are not allowed to go through the center zone
     
  • We can analyze which persons (agents) are affected by that policy
     
  • We can measure the impact of the policy on the surrounding traffic
     
  • High level estimation of traffic and emission impact

Transport policy: Limited traffic zones

Transit flow related to the ZTL

Difference after introduction of ZTL

What is the impact of the Limited Traffic Zone in the center of Paris?
 

  • Rule: Non-residents that are not performing an activity in the center of Paris are not allowed to go through the center zone
     
  • We can analyze which persons (agents) are affected by that policy
     
  • We can measure the impact of the policy on the surrounding traffic
     
  • High level estimation of traffic and emission impact

Transport policy: Parcel deliveries

A detailed study of environmental policies on parcel deliveries
 

  • Obtaining a daily synthetic parcel demand based on a synthetic population for Lyon and statistics (Gardrat)
     
  • Identifying all logistics centers in the area
     
  • Cost structures (vehicles, drivers, operational) for ICVs and BEVs (small, medium, large)
     
  • Definition of one Heterogeneous Fleet VRP per logistics center, sensitive to cost inputs
     
  • Testing of CO2 tax, ICV tax, qualitative policies

Transport policy: Parcel deliveries

A detailed study of environmental policies on parcel deliveries
 

  • Obtaining a daily synthetic parcel demand based on a synthetic population for Lyon and statistics (Gardrat)
     
  • Identifying all logistics centers in the area
     
  • Cost structures (vehicles, drivers, operational) for ICVs and BEVs (small, medium, large)
     
  • Definition of one Heterogeneous Fleet VRP per logistics center, sensitive to cost inputs
     
  • Testing of CO2 tax, ICV tax, qualitative policies

Hörl, S., Briand, Y., & Puchinger, J. (2025). Decarbonization policies for last-mile parcels: A replicable open-data case study for Lyon. Transportation Research Part D: Transport and Environment, 146, 104893.

Use cases: Community

Lille

Paris

Strasbourg

Lyon

Toulouse

Bordeaux

Nantes

Rennes

Contributors

Users

Paris

Bordeaux

Nantes

Impacts of noise on the population

  • Connection with the open-source NoiseModeling framework
     
  • Estimation of per-agent noise impact over an entire day

Le Bescond, V., Can, A., Aumond, P., & Gastineau, P. (2021). Open-source modeling chain for the dynamic assessment of road traffic noise exposure. Transportation Research Part D: Transport and Environment, 94, 102793.

Hankach, P., Le Bescond, V., Gastineau, P., Vandanjon, P.-O., Can, A., & Aumond, P. (2024). Individual-level activity-based modeling and indicators for assessing construction sites noise exposure in urban areas. Sustainable Cities and Society, 101, 105188.

Use cases: Community

Lyon

Access to green spaces and quiet areas

  • Individual-based geographic analysis of green space access per person
     
  • Recommendations on urban development

Luquezi, L. G., Le Bescond, V., Aumond, P., Gastineau, P., & Can, A. (2025). Assessing accessibility to quiet and green areas at the city scale using an agent-based transport model. Landscape and Urban Planning, 263, 105452.

Use cases: Community

Lille

Mobility pricing and park + ride

  • Implementation of a city tax to be paid when entering the city by car
     
  • Development of additional park & ride infrastructure

Diallo, A. O., Lozenguez, G., Doniec, A., & Mandiau, R. (2023). Agent-Based Approach for (Peri-)Urban Inter-Modality Policies: Application to Real Data from the Lille Metropolis. Sensors, 23(5).

Diallo, A. O., Lozenguez, G., Doniec, A., & Mandiau, R. (2025). Utility-based agent model for intermodal behaviors: A case study for urban toll in Lille. Applied Intelligence, 55(4), 282.

Use cases: Community

Lille

Lyon

Toulouse

Rennes

Shared mobility services in Rennes

  • Connection with shared mobility simulation package Starling

Manout, O., Diallo, A. O., & Gloriot, T. (2024). Implications of pricing and fleet size strategies on shared bikes and e-scooters: A case study from Lyon, France. Transportation.

Leblond, V., Desbureaux, L., & Bielecki, V. (2020). A new agent-based software for designing and optimizing emerging mobility services: Application to city of Rennes. European Transport Conference 2020, 17.

Dimensioning of shared bicycle supply in Lyon

  • Agent-based simulations on pricing and fleet sizing

Use cases: Community

Communication: Interface development

TERRITORIA price 2024


 

with Paris Saclay

  • Ambition to provide project results through interactive interfaces

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Transport modeling chain

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Replicability?

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Replicability?

Yes, eqasim-synpop for France and a handful of other cases.

Working on generalizing the methodology.

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Replicability?

Yes, eqasim-synpop for France and a handful of other cases.

Working on generalizing the methodology.

Partly, eqasim-java is accessible. Goal to publish a calibrated fully replicable simulation for Île-de-France in the coming months.

Raw data

Synthetic population

Agent-based transport simulation

Use cases

Results

Replicability?

Yes, eqasim-synpop for France and a handful of other cases.

Working on generalizing the methodology.

Partly, eqasim-java is accessible. Goal to publish a calibrated fully replicable simulation for Île-de-France in the coming months.

Using the new baseline simulation, our goal is to publish upcoming studies in a fully replicable way.

Outlook: Surrogate modeling

Surrogate models

  • Idea: Approximate the results of heavy, large-scale agent-based models
  • Evaluate policies, service configurations, ... within seconds
  • Various approaches are available - can they be refined for the transportation context?

Detailed ABMS

Surrogate model

Inputs

Outputs (KPIs)

Learning

Hours, days, ...

Seconds!

Outlook: Surrogate modeling

Natterer, E. S., Rao, S. R., Tejada Lapuerta, A., Engelhardt, R., Hörl, S., & Bogenberger, K. (2025). Machine learning surrogates for agent-based models in transportation policy analysis. Transportation Research Part C: Emerging Technologies, 180, 105360.

Outlook: Surrogate modeling

Detailed ABMS

Surrogate model

Inputs

Outputs (KPIs)

Learning

Hours, days, ...

Seconds!

A standardized protocol to set up surrogates for agent-based transport models

  • What are the relevant inputs / outputs?
  • Which surrogate architectures work best?
  • Which learning algorithms work best? (efficiency vs. solution quality)

Outlook: Surrogate modeling

Example: On-demand mobility service design

Automated generation of synthetic populations in France

Automated calibration of baseline simulations

Simulation components for on-demand mobility services

Generate a large database of ABMS
  (1) in different territories
  (2) with different operating areas

  (3) with different fleet sizes
  (4) with different service levels

  (5) ...

Outlook: Surrogate modeling

Example: On-demand mobility service design

Learning

Surrogate model

Service configuration

KPIs

Reinforcement learning

Optimal configuration

* with design constraints

Service design agent

Outlook: Surrogate modeling

Example: On-demand mobility service design

A map of optimal deployment strategies for on-demand mobility services across France

Larger scientific context

Replicability and robustness

  • Agent-based models required to assess modern technologies and policies
  • Limited confidence due to the lack of standardized evaluation and validation mechanisms
  • First step is to provide full transparency and replicability

Connecting agent-based models and surrogate approaches

  • Not to replace agent-based models (ability to explore counter-factual cases)
  • Allow exploration of larger variety of use cases and policy / service optimziation
  • Increase practical planning relevance of agent-based models

Thank you!

sebastian.horl@irt-systemx.fr

Icons throughout the presentation: https://fontawesome.com

https://sebhoerl.github.io/

Website

Contact