Towards reproducible agent-based simulations of the transportation system

Sebastian Hörl

27 May 2025

Habilitation à diriger des recherches

Habilitation à diriger les recherches

Title in French

Vers des simulations multi-agent reproductibles des systèmes de transport

Reporting jury members (rapporteurs)

• Cyril FONLUPT, HDR, Professur des universités, Université du Littoral - Côte d'Opale
• Chirine GHEDIRA, HDR, Professure des universités, IEA Lyon School of Management
• Francisco PEREIRA, Professor, Danish Technical University

Examining jury members (examinateurs)

• Flavien BALBO, HDR, Professeur des universités,ENS des Mines de Saint-Etienne
• Stéphane GALLAND, HDR, Professur des universités, Université de Technologie de Belfort Montbéliard
• Monica MENENDEZ, Professor, New York Université at Abu Dhabi
• Antonio SCIARETTA, HDR, IFP energies nouvelles
• Mahdi ZARGAYOUNA, HDR, Directeur de recherche, Université Gustave Eiffel

Date

27 May 2025

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?

Short biography

Bachelor of Science

Systems theory and control (2010 - 2014)

Short biography: Production

Publications

• 18 articles in peer-reviewed international journals
  Transportation Research: Part A, Transportation Research: Part C, ...
   
• 50 contributions to international peer-reviewed conferences
  TRB, ABMTRANS, hEART, ...

Short biography: Activities

Community activities

• Board member of the MATSim Association (since 2023)
• Program committee ABMTRANS (since 2022)
• Scientific committee TRISTAN (since 2022)
• Review for various journals (Transportation Research PartA/C/D, Transportation, ...)

Short biography: Supervision

Tarek CHOUAKI

Co-supervision of doctoral students

Simulation of on-demand services using reinforcement learning

2020 - 2023

CentraleSupélec / IRT SystemX, with Jakob PUCHINGER

6 conférence contribution

Benoît Matet

Use of mobility traces from phone data in population synthesis

2022 - 2024

Univeristé Gustave Eiffel, with Latifa OUKHELLOU & Etienne CÔME

one conférence contribution, one journal article

Jean-Giono ZEHOUNKPE

Benchmarking of population synthesis approaches

since 2024

Université Gustave Eiffel, with Latifa OUKHELLOU

one conférence contribution

Ali NAMAAN

Simulation an design of novel regional mobility services

since 2025

Université Gustave Eiffel, with Negin ALISOLTANI & Mahdi ZARGAYOUNA

one conférence contribution

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

Data

Goals

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

French population census

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

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

MOBPRO

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

MOBPRO

SIRENE

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

MOBETUD

BPE

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

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

BPE

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

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

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

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!

Munich: 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 to be presented at the 23rd International Conference on Practical applications of Agents and Multi-Agent Systems (PAAMS 2025), June 2025, Lille, France.

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!

Munich: 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 to be presented at the 23rd International Conference on Practical applications of Agents and Multi-Agent Systems (PAAMS 2025), June 2025, Lille, France.

Synthetic populations: Further topics

Logistics: Using statistics on parcel deliveries by socio-demographic attributes to generate a synthetic parcel demand data set

Hörl, S., Puchinger, J., 2023. From synthetic population to parcel demand: A modeling pipeline and case study for last-mile deliveries in Lyon. Transportation Research Procedia, TRA Lisbon 2022 Conference Proceedings Transport Research Arena (TRA Lisbon 2022),14th-17th November 2022, Lisboa, Portugal 72, 1707–1714.

Synthetic populations: Further topics

Personas: Linking synthetic populations with concepts from design science to generate future population scenarios

Gall, T., Hörl, S., Vallet, F., Yannou, B., 2023. Integrating future trends and uncertainties in urban mobility design via data-driven personas and scenarios. European Transport Research Review 15, 45.

Synthetic populations: Further topics

Charging: Deriving the electric charging demand based on person characteristics

Synthetic populations: Visualization

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?

Primary locations

Activity chains

Secondary locations

Persons

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:

Multi-modal approach

Mobility simulation

Queue-based network simulation

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

Agent-based simulation:

Co-evolutionary algorithm

Decision-making

Scoring

Home

Home

Work

Home

Time

Score

Mobility simulation

Multi-modal approach

Queue-based network simulation

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

Agent-based simulation:

Co-evolutionary algorithm

Decision-making

Scoring

Home

Home

Work

Home

Time

Score

-2

Mobility simulation

Multi-modal approach

Queue-based network simulation

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

Agent-based simulation:

Co-evolutionary algorithm

Decision-making

Scoring

Home

Home

Work

Home

Time

Score

Mutation

-2

Mobility simulation

Multi-modal approach

Queue-based network simulation

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

Agent-based simulation:

Co-evolutionary algorithm

Decision-making

Scoring

Home

Home

Work

Home

Time

Score

Mutation

-2

Mobility simulation

Multi-modal approach

Queue-based network simulation

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

Agent-based simulation:

Co-evolutionary algorithm

Decision-making

Scoring

Home

Home

Work

Home

Time

Score

Mutation

-3

Mobility simulation

Multi-modal approach

Queue-based network simulation

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

Agent-based simulation:

Co-evolutionary algorithm

Decision-making

Scoring

Home

Home

Work

Home

Time

Score

Mutation

-3

Selection

/

-2

?

?

Mobility simulation

Multi-modal approach

Queue-based network simulation

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

Agent-based simulation:

Scoring-based decision-making

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

Mobility simulation

Mutation / Selection

Simulation stabilized?

Agent-based simulation: Discrete choice

Felix Becker, Institute for Transport Planning and Systems, ETH Zurich.

Agent-based simulation: Discrete choice integration

Mobility simulation

Mutation / Selection

Simulation stabilized?

Scoring-based decision-making

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

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 making use of discrete choice models

Agent-based simulation: Calibration

Parameters

• Capacity factor by OSM category
• Speed factor by OSM category
• Specific factors for Paris
• Alterantive-specific choice constants

Objectives

• Global mode shares
• Mode shares by Euclidean distance
• Travel time distributions
• Point-to-point travel times

Work in progress

Goal: Publish a well-calibrated openly accessible simulation of Île-de-France

Agent-based simulation: Further topics

Resource constraints: Synchronization over limited resources like

• vehicles in a household / peer group
• parking spots
• charging stations

Hörl, S., Sobieraj, J., Axer, S., Rewald, H., 2023. Resource-constrained replanning in MATSim applied to the simulation of peer-to-peer car sharing services. Procedia Computer Science, The 14th International Conference on Ambient Systems, Networks and Technologies Networks (ANT 2022) and The 6th International Conference on Emerging Data and Industry 4.0 (EDI40) 220, 698–703.

Agent-based simulation: Outlook

Surrogate modeling

• Wrapping heavy, computational expensive agent-based transport simulations with surrogate models
• Quick approximation of their outcomes to enable evaluation of large scenario analyses

Natterer, E., Engelhardt, R., Hörl, S., Bogenberger, K., 2025. Machine Learning Surrogates for Optimizing Transportation Policies with Agent-Based Models, in: 12th Triennial Symposium on Transportation Analysis (TRISTAN XII). Okinawa, Japan.

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

• 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

Use cases: On-demand mobility

On-demand mobility: Integration with public transport

• How to combine on-demand mobility systems with public transport?
   
• Interesting context: Grand Paris Express
   

On-demand mobility: Integration with public transport

• How to combine on-demand mobility systems with public transport?
   
• Interesting context: Grand Paris Express
   
• Trying to maximize the complementarity between mass transit and on-demand transit

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.

On-demand mobility: Extensions

• 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: Outlook

Algorithmic fairness

• 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.

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.

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!

https://slides.com/sebastianhorl/hdr

sebastian.horl@irt-systemx.fr

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

Tarek, Jakob, Tjark, Arthur, Yann and + at IRT SystemX

... my former colleagues at ETH Zurich (and especially Milos)

... our collaborators at Volkswagen

... and everybody who has contributed in one way or another!

... Latifa and Mahdi at Université Gustave Eiffel

Agent-based simulation: Calibration

Decision-making

Mobility simulation

Synthetic demand

Mode shares

Agent-based simulation: Calibration

Decision-making

Mobility simulation

Synthetic demand

Mode shares

Objective calculator

Reference data

Agent-based simulation: Calibration

Decision-making

Mobility simulation

Synthetic demand

Optimization algorithm

Behavioral
parameters

Mode shares

Objective calculator

Reference data

Agent-based simulation: Calibration

Decision-making

Mobility simulation

Synthetic demand

Optimization algorithm

Behavioral
parameters

Mode shares

Objective calculator

Reference data

SPSA, CMA-ES, ...

Agent-based simulation: Calibration

Decision-making

Mobility simulation

Synthetic demand

Optimization algorithm

Network
parameters

Travel times

Objective calculator

Reference data

SPSA, CMA-ES, ...

Agent-based simulation: Calibration

Decision-making

Mobility simulation

Synthetic demand

Optimization algorithm

Network
parameters

Behavioral
parameters

Objective calculator