FixT : Fixing the traffic in Zurich

Sooo traffic is a huge problem for many cities… afterall who never experienced being stuck in the car during the rush hours, spending 40 minutes of valuable time for a trip that normally takes 10.

Not to speak of the wasted fuel.

Common sights in a city

According to ourworldindata road transport accounts for 15% of the global CO2 emissions, CNN claims that the average commuter like you and me wastes 54 hours a year due to congestion.

Governments around the world are ready to invest to solve this problem, in 2021 the US alone invested 1.2 Trillion $ in the Infrasturcture Bill and according to Juniper the smart traffic management market was valued 10,423.7 million in 2022 and projected to grow of an additional 13.8% before 2030.

This investments show the attempts of the governments to meet agreements such as the Paris Agreement in order to deal with the current climite crisis. Moreover investements on this sector help to mitigate the mind-boggling economic toll of traffic.

So how do we solve this?

The concept of using information like traffic density in order to control traffic is really not new. Many cities, like for example Zürich,Switzerland , are already implementing smart traffic management systems to mitigate congestion.

For a long time the controllers used in this field have been model based, meaning that model estimation techniques were applied to data extracted from a city to obtain an approximated description of how the traffic behaves. Once this information was obtained, then the system could be controlled using techinques such as Perimeter Control. Famous examples of Model Predictive Controls (MPCs) are the ones devised by Professor Nikolaos Gerolinimis at EPFL.

This controllers work great in both in congested and uncongested settings, however they come with an obvious drawback, the model itself. This is because model estimation can be extremely expensive for large system such as a city.

It would be great if we could have a way of controlling traffic in cities directly from data, in a end-to-end data-driven fashion!

This approaches would drammatically reduce the cost of implementing control into a city. Together with the fact that companies like DataFromSky are lowering the cost of infrastructure for extracting traffic data, this approaches have the potential to help cities become smarter faster.

This is exactly what this project is about, during the course of six months at ETH Zürich, me, Carlo Cenedese and Alberto Padoan investigated the feasibility and effectiveness of such methods.

How did we do this?

There are several data-driven algorithms that could be used for this porpuse, there exists also a data-driven formualtion of MPC! One very interesting idea would have been to use Deep Reinforcement Learning to solve this, however the implementation of this systems in the real world requires safety guarantees, the why is obvious, urban systems are strategic hubs in which stability and safety must be ensuerd. DeeP RL, as all Deep Learning algorithms, has a problem of interpretability. Moreover devising theoretically the safety guarantees and feasible settings for this algorithms is a real challenge. Therefore we settled on using Data-Enabled Predictive Control (DeePC), which is a data-driven control algorithm first devised in 2019 at the Automatic Control Laboratory by Jeremy Coulson.

Obviously you cannot test these algorithms on the real city, luckly for us the ETH spinoff company Transcality developed a highly detailed simulation of the city of Zürich, the simulation is embedded in the state-of-the-art traffic simulation software SUMO.

net net The Bellevue intersection

We tested DeePC against a no control baseline policy and a linear model predictive control formulation devised by Nikolaos Gerolinimis. The algorithms were tested in two different congestion conditions shown below

net An uncongested day net A very congested day

How did it go?

DeePC worked great in the simulation setting, it was able to beat both the no control baseline and MPC, it achieved a reduction of 13,0% in travel time and 8,2% reduction in CO2 emissions, saving 65 Tons of CO2 in just one demand peak. For more information on the methodology and results, we presented our findings in a paper submitted at the European Control Conference 2024 (at the time being waiting for peer-review).

As a byproduct of this study the first framework for benchmarking and automated testing of traffic algorithms has been born. Completely written in Python, it allows the user to simulate, test and debug control algorithms without the need of directely interacting with SUMO, it will be soon released as an open source repository control_sumo.

control_sumo

Conclusion

I always feel very acomplished when i finish a project, in particular when it has the potential of being impactful on society, in FixT i found a great mix of complexity, challenge and curiosity. I learned a lot of great data analysis and optimal control concepts during this time, the project is not really over however! As the repo becomes open source many iterations and expansions will be carried out.

Sources

  • Hannah Ritchie (2020) - “Cars, planes, trains: where do CO2 emissions from transport come from?” Published online at OurWorldInData.org. Retrieved from: ‘https://ourworldindata.org/co2-emissions-from-transport’

  • https://www.juniperresearch.com/researchstore/sustainability-technology-iot/smart-traffic-management-research-report?ch=smart%20traffic%20market

  • https://www.bloomberg.com/news/articles/2018-02-07/new-study-of-global-traffic-reveals-that-traffic-is-bad

  • https://arxiv.org/pdf/1906.04679.pdf

  • https://arxiv.org/abs/1811.05890

  • https://transcality.com/

  • https://sumo.dlr.de/docs/index.html