Thank you, Peter.
Yeah. So today I want to talk a little bit about Matzim.
Matzim is a transport simulation software that is being used at SBB, the Swiss Federal
Airways, but it is actually an open source tool that has been around for quite some time.
So obviously there is also a little bit talk about Matzim itself and the software and what
it does and how you could use it if you are ever interested in that.
That's on the agenda, so I'll briefly explain what Matzim does and why we find that useful
at SBB.
And we also contribute actively into the Matzim code and I'll give some examples of that.
And yeah, since you might wonder why on earth are you even bothering, I'll give you some
examples of our work with the software.
So what is Matzim and why is it useful?
If you have that elevator speech moment where you have to explain your work to your CEO
and then they ask you what you're doing, then I tend to say I'm playing SimCity, but
with complex econometric data behind it, so you have all these weird formulas somewhere
in there.
Then the elevator area is over and I have more or less explained what I'm doing and
then the CEO knows that we have some guy playing SimCity all day.
Well, there's a bit more behind it, but in brief that's what you're doing.
So we are simulating transport and we simulate people's behavior using transport during the
day.
Matzim stands for Multi-Agent Transport Simulation and it has been around for roughly 20 years.
It started as a purely academic project between ETH Zurich and TU Berlin.
On a side note, that also explains why a Berlin guy is now living in Switzerland.
So you can kind of imagine my background, but it has evolved over the years and there
are many models around the world and quite a few of them are actually fully built on
open data and are publicly available, not ours for some reasons, but for example there's
quite a scenario about Berlin that you can download and you can see the data where it
comes from and you can start playing with the model.
Whether this is useful for anyone, I can't say, but I guess I think it's useful for some,
even mostly PhD students to be fair.
There are commercial users around the globe as well.
Among us, the SBB, there's Volkswagen who have quite a strong development, also into
the Matzim core, but they're not as open to talk about that as we are probably.
Then there are models in Melbourne, there's one at the Berlin Transit Agency, so it has
some standing right there.
There's a book, there's code, there's a license and for the last couple of weeks there's also
an association that kind of brings the whole thing together.
Now, how does it work?
So imagine you have a lot of data.
You have census data, for example, you register data, you know where people live in a city
or you just make that up, you replace people somewhere.
You have econometric data that is value of times.
You know what a person's intention is.
If they travel by train, then the value of time is maybe six euros per hour and if they
go by car, then it's maybe 10 euros per hour or the other way around.
You have a road network that can come, for example, from OpenStreetMap, that is a very
typical use case.
You have a timetable for public transport, typically GTFS.
You have count data.
Many of the topics discussed in the previous talks are actually input data for us and that
is a lot of input data.
What we do then is we add some generic algorithms that basically randomly decide to tell people
during the day, change your route, change your transport mode when you go from one activity
to another or change your departure time choice and then we let that run the same day.
It's a bit like Groundhog Day, 200 times, 500 times and mix people up and let them try
out new things.
This is what we call the Matsum loop that is also somewhere on my T-shirt.
What comes out of it is actually output data, even more of it.
You have individual daily plans.
You know what your synthetic population is doing during the day, where they go shopping,
what transport modes they use.
You have mode choice for each strip, whether people tend to take car to get from A to B
or public transport, depending on what's the offer.
You have time-sharp traffic loads, so a lot of data to analyze and do your policy planning.
You have distances, you have all kind of aggregate data that you can then use and play with.
Obviously, this calibration process for a model that it really depicts the real world
in an initial stage is kind of what's the long story behind model building.
What can you use the whole thing for?
Of course, transport policy evaluation, what happens if there's a new road, what happens
if there's a new railway line, what happens if there's a new price.
You can do a person specific.
You know who's affected by transport policy because you have this agent-based paradigm
behind it.
You can also calculate, for example, accessibility.
A lot of things are actually happening where MADSIM is being used when it comes to on-demand
transport modes.
You can really do your fleet scheduling, your fleet planning.
You can say, okay, what happens if we have a lot of automated vehicles that replace passenger
cars and what's the advantage of that?
All these kind of future scenarios you can use MADSIM for, well, basically playing SimCity.
The MADSIM project has been around for almost 20 years and historically it has been administered
by the universities, so ETH Zurich and Tew Berlin.
Professor's grow older and at one point they retire and the person that comes next is maybe
not as interested in such transportation simulation anymore.
Since last year, the whole MADSIM project is built on an association level, so that
there's also some funding from other users to maintain build servers and all that stuff.
The association also organizes things like the user meeting that is held annually or keeps
track of all kind of developments, publish a newsletter, all these kind of stuff.
Now at the Swiss Federal Railways, how did we start with that?
It's a very brief timeline on one slide, but I think it's kind of interesting.
In 2016, our CEO saw a presentation about MADSIM and decided we need this at SBB, please
buy MADSIM.
Well, as it happens with open source software models and open source software at all, buying
the whole thing wasn't as easy and the whole procurement process didn't quite work out,
so the task was delegated to the department that deals with classical transport models,
so it's actually part of somewhere in the passenger division, this is also where I'm
working.
It came up with some challenges, for example you needed someone who knows programming Java
and those people didn't exist in that department, but things that can be overcome and you need
like proper computers actually because if you want to run proper big models then having
a nice tiny laptop isn't sufficient.
That was also something to overcome, also thanks to the IT at Peter's department in
the end.
At least we didn't kill it.
At least they didn't.
Yeah, but the whole thing, building a model for Switzerland in MADSIM took three years
and ran from 2017 to 2020 and along the way we noticed that there are several additions
into the code base that need to be made to make this a useful project for us and then
at one point you see okay we need to decide, do we commit this back into the MADSIM core
or do we keep that in our secret chamber and luckily people chose widely and this is
all open source and this was actually a management-backed decision so I'm very happy with that.
So in 2018 the first release of the model that I will present in a moment was released
and since 2020 we have an annual release cycle of a new, of a transport model that is multi-model
and MADSIM based in Switzerland and that can be used for all kind of policy studies.
Our contributions to MADSIM I just want to showcase two.
First of all one that is called oddly the Swiss rail rector so it's an rector based public
transit router that works really fast because if you want to route millions of people within
a reasonable time frame then this is what you need and compared to what we had before
it was like many, many times faster and the whole simulation was actually sped up by factor
three and what is also important to say, the Swiss rail rector is, well it's a Java package
but you don't have to, it's tied to MADSIM, it uses MADSIM data structures but you don't
actually have to use it for MADSIM problems so in a way it's something that can be used
instead of OpenTrip planner but OpenTrip planner has other advantages but if you really need
to route a lot of routes at the same time then it's something to look at.
And then we, yeah now we have a bit of a fancy routing algorithm so it also knows like range
queries, queries and intermodal access and egress and the last point is very, very important
because if you want to model stations you really need to have an idea how many people
arrive by food, how many arrive by bike and so on and that is not an easy question to
answer and apart from the routing problem which is already quite complex it's also a part of the,
you don't have that much empirical data about it so, but it's really one of the most useful
features by now from our model so we're happy to have that and yeah as I said you can use that
kind of independent of the rest of MADSIM so it's well worth checking out.
Then there's another contribution where I was a bit deeply involved, it's basically,
there is a traffic flow simulation in MADSIM that is typically queue based, I didn't go
into detail here because I don't have enough time for that but we replaced this by something
that is just like roughly two times faster for the whole simulation process and it's called
Hermes because well it can fly now but has less plugability so depending on the use case of your
simulation you can use one or the other and they're kind of interoperable and yeah this brings down
simulation run times for Switzerland scenario so both the router and Hermes to something like 24
hours and since these are typically AWS instances it actually saves a lot of money to have models
that run reasonably fast and for calibration process you maybe need 50 of those runs so it's
it's kind of sums up yeah now what we use MADSIM for first of all there's a model called Zimba Mobi
this is where I'm the product owner so I know a lot about it and it literally depicts the everyday
mobility of eight and a half million people in Switzerland so basically the whole population
it includes all major public transport modes so walking cycling taking the bus taking the car
and it has a representation of the transit schedule obviously and also the whole road network
and since it uses MADSIM the whole people's behavior or agents behavior is microscopic
so including first and last mile decisions and hope that video works now
yeah it does so right now imagine it's 8 a.m in Switzerland and you see people
in those blue dots being at home and these light blue dogs people starting their work time
and now we zoom in into a region somewhere in Eurik and see what people do there so to get from
one place to another they need to travel obviously and they could travel by car that then there are
in those little gray boxes or they could take the train or public transport then there are in
those little red boxes and they get from one place to another and obviously you can run your
analysis on that some public transport vehicles are maybe more crowded some are less crowded
and yeah you can sum that up over the day and see what's going on
and now we zoom in again
and what you can do is see who's alighting at certain stations what kind of passenger groups
are they are there do they have a regional subscription do they have a half fare card or
do they have ordinary tickets for example and also on the highway you can see during the day
how many people are currently on their way to work how many people are on their way home
and who's just a truck and who's doing other things and this is all in the model and you can
like analyze a lot of things obviously we are a bit more tied to public transport related analysis so
station access and egress is different at certain points so in Jettikon for example there's more
people arriving by bus to the station than in Aarau or if you have this city in Baden here
then you can see people that reach the stations from nearby typically walk and people that come
from further field further field they take the bus and these kind of analysis is really useful
for example for station design and station planning
and yeah so typically use cases for the model are then also the development of rail lines and
designing of stop locations and time and the effects of timetables so
it's nice that you created a nice timetable but who's going to use it and how many people will
be on the trains this is an answer we can give and we can analyze what's happening around the
stations so that was just the video that I showed and we can also see what's the effect of certain
land use policies for example so we don't only have a model that depicts today but also one for
2030, 2040, 2050 so we kind of know how according to today's assumptions Switzerland is going
to evolve and then we can do policy planning with this yeah and these are kind of future scenarios
yeah and just one example for example over the next 20 years there will be roughly
20 to 30 new railway stations opening up mostly along existing lines and
and very often these stations are being built because there's something happening around them so
there's a new development coming new housing or a new commercial area being built or something like
that and we can just like in Simcity we can add those little houses into the model and add people
there give them daily plans and for example this is now in the city of Sankt Gallen where not a new
station is being planned but the the moving of one station to another place so basically that
station goes from left to the right and then lots of houses are being built there and with the tool
we can say okay what's the beforehand we had at both those stations there are roughly 4 000
passengers a day and now it's roughly 6 000 so that would be the effect of the policy of things
happening there and and these numbers help you to dimension those stops properly
yeah another application that doesn't come from my department so please don't ask me
doesn't question about it but i think it's interesting enough to be presented here is
that we also want to go into deep down knowing what's happening along the railway corridors so
Matsim has a mobility simulation i talked about that earlier and my colleagues decided okay we
can replace this with something that we call Ray Sim and that actually has tracks and signals in there
and blocks and we can start playing around with that and and do roughly also capacity planning
and on a much easier level than it is usually being done so that you still don't need
need to have an idea of every signal that's on the tracks but and of every switch but
of the whole idea roughly you need to know whether a track is single track or double track for example
and so the outcome of this is now also a little video you have two trains
um
one coming from down here it has currently a speed of six six meters per second but it
wants to go it's currently accelerating to 11 meters per second and there you have also a train
that is six meters per second and wants to accelerate to 14 meters per second and this train
wants to go this way and this train wants to go that way and there's a station and they both
interfere at one point so obviously we don't want them to crash so as you can see the train coming
and the red lines are basically the blocks that are being blocked in front of the train so depending
how how fast the train is the train is faster than the braking distance would be longer so
more blocks are being blocked so um and then you can see the train that comes from the right
has the right of way and and and the left train got a red signal and is breaking down
and now that right train is passing switch goes to green and the other train can enter the station
well yeah um and obviously you can connect that with the rest of matzin so you know
how many passengers are on the train and then you can do your policy planning again
if there's um if there's a heavily delayed freight train that but that would generate
you that and that amount of money maybe you want to accelerate it but then you see oh no it interferes
with all of our daily commuters they would be very angry so you can do your policy planning
around this it's still in an early stage the that microscopic railway simulation but ultimately
this is where we want to go to it's also released as as a matzin contribution called rail sim so it's
part of the matzin code everyone can use it and i think it's a way to go in that direction as well
but please don't ask me too many questions about that but i can connect to people who know about it
so to wrap up
matzin has helped sbb massively to to understand customer behavior and
committing to open source in our point of view has really paid off here and it's also the way to go
for us um yeah and these models are very very complex but um that is what they are no matter
whether you use commercial or open source software um oh yeah that one has to come too um and uh yeah
if you want to know more about matzin um there's an annual meeting um this year it will be in
um as part of the heart conference it's a transport conference um on the 17th of june in
in outer university so yeah hey thank you um yeah and um i'm happy for questions
even though i only have five minutes
thanks a lot for the presentation uh quite a bit of the the transport systems have an historical
background for example probably some of them are based on the industrial need back at like
20 years ago or some of them are related to new developments in the city and some of these things
are also presented in fact at open street map in these kind of resources are you able to extract
amenities maybe or an historical sense or is the distributioner and the social demographic
something that you get as a matrix how do you get this kind of distribution and sorry second
yep do you import EFS and these other stuff um yeah both good questions so the first one um so
we do have the census data from the federal office of statistics in switzerland and they also have an
idea how it looks like in the future so we are in a very lucky situation there's a lot of data
available and also publicly um in fact there's also a transport model available for switzerland
publicly but unfortunately with a closed source license where you need a software that costs you
roughly uh 10 000 euros a year so that doesn't help a lot um so we rather build our own models
and the other question is um yeah typically we uh one would use gtfs as an as the main data source
for public transport data since we are the railway operator and we have time tables in all kind of
formats we use a different one but if you were to build a model for matzim typically you would use
opus creed map and and and gtfs um yeah
yep you um what for the model of the swiss transport uh how do you do that there's
three three million agents that are being simulated eight and a half yeah eight and a half
so what's what's the typical like how how much can it scale like how many agents can you have in one
simulation um yeah that um so the the models do scale um but um there's there's an upper limit
to what is useful because um so switzerland is still a useful also in terms of a regional
scope because there you have many long distance commuters um but if you are using matzim to for
example for for really long distance um choices um then you would simply remove everyday commuters
so i also in a previous life i created a model for sweden that also worked but it's also roughly the
same number of people um and but there are simulations for for cities in china which have 20 million
people um and what you can do is you can cut down the number of agents that you simulate and and
increase network capacities and so on so there there are ways to deal with that
yeah so you mentioned that you can feed open street uh uh map data into matzim
but does matzim also provide tools to add new assets or population models how
there are tools that allow the addition of of new people if you don't want to hard code it
some of them are commercial i mean there are some spin-off companies around matzim who provide
this as a service but you can also do it on your own it's just basic java or python code that you
can use for that so um yeah and if you do transport modeling um for public transport you
would probably rather edit the gtfs than editing the matzim schedule
so i think there are a lot of interesting yeah i'm here so sorry
yeah well i think we one short question one question is possible yeah one more
yeah i don't know how do you determine the accuracy of your model ah oh yeah that that is another talk
of of an hour um so getting getting getting models right and calibrating them properly um is um
is mostly countator that you need and in switzerland we have something that's called the microsensors
that where where where we ask people every five years about their mobility behavior and that is very
very accurate and has a lot of data where that is useful to calibrate models but um
it's always a fair question if someone presents you a transport model to ask how is it calibrated
okay so
you