Maryland Integrated Land-use and Everyday-travel Simulator

From census records
to cars on the road

MILES is an agent-based modelling platform for Maryland. It synthesises every household of the region, moves it through the housing and job markets year by year, plans each person's weekday and simulates the resulting traffic on the road and transit networks.

Background: simulated cars of a 2024 weekday in the Baltimore region, 04:00 to midnight, from the MILES MATSim run (10 % sample). Each car follows its simulated route, departure time and travel time.

12.5 Msynthetic persons
1,588zones in six states
27estimated choice models
3,946activity distributions
2,308count stations tested

The platform

Five models, one chain

Each stage hands its output to the next. Land use runs year by year from 2016 to 2024; activities and traffic are simulated for a weekday of 2024. Select a stage.

Synthetic population

A conditional variational autoencoder generates households and persons of 2016, including combinations of attributes that the sample does not contain, then places them in zones and dwellings and gives workers a job.

Engine
Conditional VAE, Python and PyTorch
Trained on
ACS PUMS 2012–16, 253,029 households in 96 PUMAs
Produces
Households, persons, dwellings, jobs and workplaces of 2016

Land use

SILO moves the population through time: births, deaths, marriages, residential moves, job changes and car ownership, with MILES's estimated residential-location, workplace and car-ownership models.

Engine
SILO, Java, yearly steps 2016 → 2024
Sensitive to
Car times and accessibility, housing costs, vacancies, jobs, incomes
Produces
Population, dwellings and jobs of every year

School location

Every student is assigned a school in each SILO year with three multinomial logits, respecting attendance boundaries, enrolment and capacity.

Engine
Inside SILO, every simulated year
Models
K–8, high school, college
Produces
The school of every student

Activities: MAST

Each person receives a full weekday: activities drawn from survey distributions, scheduled, located with estimated logit models, and tour modes chosen with household car allocation and joint tours.

Engine
MAST, Java, built on the ABIT code base
Estimated on
2017/18 Regional Travel Survey, estimation half
Produces
One weekday plan per person: activities, locations, times, modes

Traffic

MATSim routes every trip and shifts departure times on an OpenStreetMap network with the region's transit timetables and values of time by income.

Engine
MATSim 2024, Java
Network
OpenStreetMap roads 2024; GTFS transit
Produces
Routes, link volumes by hour, travel times, boardings

Why MILES

Built for policy questions at the level of the person

Every person is an agent

Households, persons, dwellings, jobs and schools are simulated individually, so results split by income, age or place of residence.

Deep generative population

A conditional VAE generates households beyond the combinations seen in the sample, conditioned on PUMA and household size.

Land use that evolves

Moves, job changes, births, deaths, marriages and car ownership update the population every year from 2016 to 2024.

A full weekday for everyone

Activities, schedules, locations and tour modes, with household car allocation and joint tours.

Dynamic traffic

Every trip routed in MATSim with departure-time shifts, transit timetables and income-based values of time.

Estimated and tested, not tuned

27 choice models estimated with Biogeme; every stage tested on data it never saw. Counts play no role in estimation or calibration.

Validation

Tested on data it never saw

The travel survey is split into an estimation half and a hold-out half; land use is calibrated on 2017–20 and tested on 2021–24; traffic is compared with 2024 counts that enter no step of the model.

Validation results Interactive count map

Synthetic population
Held-out 10 % of ACS PUMS records; ACS 2014–18 by PUMA
Land use
Calibrated on 2017–20; tested on ACS and Census 2021–24
School location
Survey hold-out half; ACS enrolment
Activities
Survey hold-out half; ATUS, ACS and LODES 2024
Traffic
2024 counts at 2,308 mainline stations, continuous counters, screenlines

Command line

Install and run

MILES installs as a Python package with one command, miles, for every step of the chain. Java engines and data paths are set once in miles.toml.

Installation Running the chain

# get the code
$ git clone https://github.com/Tomal121186621/MILES.git
$ cd MILES && pip install -e .

# point MILES to the data and engines
$ cp miles.example.toml miles.toml
$ miles config

# run the chain, stage by stage
$ miles population all
$ miles landuse run
$ miles activity run
$ miles assignment run

Developed at

Center for MultiModal Mobility in Urban, Rural and Tribal Areas

University of Maryland, College Park