Sebastian Hörl

28 September 2026

MATSim User Meeting 2026

Benchmarking on-demand mobility algorithms using a new
remote dispatching interface for MATSim

Introduction

• MATSim provides the DRT / DVRP packages that allow simulating on-demand mobility services
   
• As standalone simulations (given fixed trips) or as part of the whole loop (agent deciding for DRT)

Fleet dispatching

• Fleet dispatching means defining which vehicle should pick up and drop off which requests and in which order


• Assignments need to adhere to constraints (latest pickup time, latest dropoff time, maximum detour, ...)
   
• DRT Algorithm
  Default implementation in MATSim
  Myopic insertion heuristic
   
• Alonso-Mora
  Overrides default implementation
  Dynamic reconstruction of schedules

matsim-org/alonso-mora

Idea

• Let's flexibilize the way we can hook dispatching algorithms into MATSim!
   
• Rapid development using Python and other languages
• Reuse of existing libraries (machine learning, inference, ...)
   
• Reuse of existing algorithms in literature
   
• Concrete use case: connecting Fleetpy to MATSim

TUM-VT/FleetPy

Messaging concept

• MATSim (server) and dispatcher (client)
  
   
• Connection via ZeroMQ
  • REQREP mode: one request follows a response follows a request
  • Keeping the message exchange simple
    
     
• Messages encoded in JSON
  • Not optimal, but easy to follow (potential to replace)
     
• Dispatching step parametrizable in MATSim, for each
  • MATSim sends state (vehicles, requests, ...)
  • Dispatcher answer with assignment (vehicle schedules, ...)

Additional information

Implemented

• Request network information to have an idea about the topology
  
  
  
  
  
  
  
  
  
  
   
• Request travel time information from MATSim
  
  
   

Future

• Routing service to load off routing information to MATSim
• Charging state, charging stations, ...
• ...

Example implementations

• Random assignment
  Random request meets random vehicle
   
• Bipartite matching
  Minimize overall empty distance
   
• Roaming dispatcher
  Cruising and serving next found request
  Manual routing using networkx
   
• Insertion dispatcher
  Python-version of DRT algorithm (almost)
   
• Q-learning dispatcher
  Learns a rebalancing matrix over iterations

Chouaki, T., Hörl, S., & Puchinger, J. (2022). Implementing reinforcement learning for on-demand vehicle rebalancing in MATSim. Procedia Computer Science, The 13th International Conference on Ambient Systems, Networks and Technologies (ANT 2022), 201, 134–141. https://doi.org/10.1016/j.procs.2022.03.020

Example implementations

• Random assignment
  Random request meets random vehicle
   
• Bipartite matching
  Minimize overall empty distance
   
• Roaming dispatcher
  Cruising and serving next found request
  Manual routing using networkx
   
• Insertion dispatcher
  Python-version of DRT algorithm (almost)
   
• Q-learning dispatcher
  Learns a rebalancing matrix over iterations

Chouaki, T., Hörl, S., & Puchinger, J. (2022). Implementing reinforcement learning for on-demand vehicle rebalancing in MATSim. Procedia Computer Science, The 13th International Conference on Ambient Systems, Networks and Technologies (ANT 2022), 201, 134–141. https://doi.org/10.1016/j.procs.2022.03.020

Example implementations

• Random assignment
  Random request meets random vehicle
   
• Bipartite matching
  Minimize overall empty distance
   
• Roaming dispatcher
  Cruising and serving next found request
  Manual routing using networkx
   
• Insertion dispatcher
  Python-version of DRT algorithm (almost)
   
• Q-learning dispatcher
  Learns a rebalancing matrix over iterations

Chouaki, T., Hörl, S., & Puchinger, J. (2022). Implementing reinforcement learning for on-demand vehicle rebalancing in MATSim. Procedia Computer Science, The 13th International Conference on Ambient Systems, Networks and Technologies (ANT 2022), 201, 134–141. https://doi.org/10.1016/j.procs.2022.03.020

Benchmarking testbed

• Road network of Paris provided as an example (OpenStreetMap)
   
• Fleet generation script: random distribution of vehicles in the network
   
• Demand generation script
  • Definition of N attractors / emitters with morning / evening factors
  • Requests are generated from / to those points depending on time of day
  • Represents directionality by time of day and spatial imbalance

Benchmarking results

Three dispatchers

• Euclidean bipartite matching
• Insertion heuristic (DRT)
• Q-Learning

Analysis

• Average over 10 seeds

Benchmarking results

Three dispatchers

• Euclidean bipartite matching
• Insertion heuristic (DRT)
• Q-Learning

Analysis

• Average over 10 seeds
• Q-Learning: 25 iterations

How to run

As a module

• Replaces individual DRT modes (configurable)
• Port given in configuration or written to tmp

sebhoerl/matsim-remote-drt

Teaser: LLM-based algorithm adaptation

• Algorithms are often general purpose
   
• What if we integrate context-specific heuristics (bridges, bottlenecks)?
  • Local knowledge > translation into rules > implementation > testing
  • Work-intensive process!
     
• Let's let an LLM take an algorithm and customize it iteratively!

Thank you!

sebastian.horl@irt-systemx.fr

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

https://sebhoerl.github.io/

Website

Contact