• none
Computer Science XI - Modeling and Simulation

Predicting Trip Destinations with Neural Networks

General Information

Supervisor Leo Strobel
Language German or English
Level Master


Realistic mobility demand is an essential input for various scientific simulations. For example, we can model the spread of diseases better if we can determine what groups of people come into regular contact. Other examples are the performance evaluation of close-range peer-to-peer communication systems and transportation system planning. However, for privacy reasons, it is not desirable to track the movement of the entire population of a country to acquire the necessary data. Instead, it is better to create synthetic mobility demand with an agent-based simulation that is representative of the original population without intruding on the privacy of individuals.

This is the purpose of the open-source mobility demand generator OMOD we developed at our chair (

A crucial step in creating synthetic mobility demand is determining where people will conduct activities like shopping or working. In classical models, this is done in two steps. First, the number of trips that originate and end in all locations is determined with linear models. Then, a so-called gravity model connects origins and destinations (See Modelling Transport, Ortúzar and Willumsen).

In OMOD, we use a similar approach adapted to fit the agent-based viewpoint. However, this approach runs into limitations if non-linear features are to be included for the attractiveness of a location. For example, there is a complex relationship in clusters of shops. On the one hand, a group of shops might be more attractive than the sum of its parts because a person could shop for multiple things at once there; think of a city center. On the other hand, at some point, a location already provides a shop for most things, and shops start to compete, reducing the marginal attractiveness of additional shops.

Another problem that arises from the classical formulation is finding a suitable calibration for the parameters of the function. This is usually done using the maximum likelihood method. However, solving the resulting optimization is increasingly difficult with a rising number of parameters. Therefore the methodology is limited in its ability to model more complex relationships between mapping information and sociodemographic-features of individuals.

The goal of this work is to find alternatives to the classical implementation of the destination choice using modern machine learning techniques, such as neural networks. The techniques can be applied to directly determine the destination choice or, alternatively, to find suitable parameters for the old implementation. 



  • Confident with Python


  • Understanding the Maximum-Likelihood concept
  • Experience with a machine learning library
  • Experience with handling data tables (>= 100 k rows)