17.02.2026

What is reinforcement learning? – AI learns from urban trial and error

Building design

City

Aerial view of a German city taken by Chundy Tanz, showing modern urban architecture and a clear street layout.

Artificial intelligence that learns from the city’s mistakes? This is exactly what reinforcement learning is all about – a method that not only analyzes urban space, but actively helps to shape it. What sounds like futuristic science fiction has long since become reality and could fundamentally change urban planning. If you want to know how algorithms deal with urban trial-and-error processes, you’d better read on – because this is where it will be decided how cities will be built, designed and even governed in the future.

Definition and functionality of reinforcement learning in the context of urban systems
Application examples for reinforcement learning in urban planning, traffic control and environmental management
Opportunities and challenges in the use of AI-supported, self-learning behavior in German cities
Technical requirements, data sources and ethical issues surrounding reinforcement learning
Differences to other AI approaches such as supervised and unsupervised learning
Innovative projects from Europe and around the world with a role model function for German-speaking countries
Risks such as algorithmic biases, governance problems and transparency deficits
Potential for new, dynamic and resilient urban planning with adaptive systems
Specific recommendations for municipalities to get started with reinforcement learning applications

Reinforcement learning: what is behind the learning algorithm?

Reinforcement learning is a branch of artificial intelligence that differs fundamentally from traditional machine learning. While supervised learning involves feeding an algorithm with many examples and the corresponding solutions, and unsupervised learning involves discovering patterns in data, reinforcement learning works according to the principle of trial and error. The algorithm – usually referred to here as an agent – moves around an environment, tries out actions and receives an evaluation, a so-called reward or punishment, for each action. The aim is to develop a strategy that achieves the highest total reward in the long term.

In an urban context, this means that an algorithm could, for example, learn to optimally direct traffic flows by trying out different traffic light circuits, receiving feedback on traffic jams, waiting times or emissions and gradually improving its action strategy as a result. The environment here is the complex urban system, in which many factors such as weather, roadworks, school start times or major events interact. Unlike fixed rules, reinforcement learning can react flexibly to new situations because the algorithm continuously learns and adapts its strategy independently.

The trick is that reinforcement learning is particularly suitable for problems where there is no simple, clear solution and the system relies on feedback in order to improve. This is precisely the case in urban planning. The complexity of cities is almost impossible to fully understand. They are constantly changing, full of uncertainties and surprises. Reinforcement learning not only accepts these uncertainties, but also uses them as a learning field. Every unexpected development becomes valuable experience for the algorithm, which ideally plans not only for the short term, but also sustainably and resiliently.

That sounds like a revolution – and it is, to some extent. But implementation is not quite that simple. Reinforcement learning is a data-hungry process: The more information the system receives, the better it can make decisions. This requires good sensor technology, powerful data platforms and a clear target definition. Without these foundations, any AI approach in urban planning will quickly come to nothing or produce results that have little to do with urban reality.

However, there is another special feature: reinforcement learning can not only handle traditional data such as traffic figures or weather data, but also qualitative feedback, for example from public participation processes or social networks. This opens up completely new possibilities for mapping urban complexity and integrating it into planning processes. Suddenly, the city becomes a laboratory for learning machines – and the planner becomes a conductor in the orchestra of algorithms.

Fields of application: How reinforcement learning makes cities smarter

Probably the best-known application of reinforcement learning in urban areas is adaptive traffic control. In many large cities, traffic lights are already equipped with self-learning algorithms that are able to analyze traffic data in real time and optimize their switching times accordingly. The aim is to avoid traffic jams, improve air quality and shorten journey times. Cities such as Los Angeles, Zurich and Singapore have been experimenting with such systems for some time – and the results speak for themselves. In Zurich, for example, a pilot project reduced the average waiting time on main traffic arteries by over 15 percent. This was made possible because the algorithm does not simply switch according to fixed patterns, but evaluates the impact of each decision and learns from it.

However, the potential of reinforcement learning is far from exhausted. In the energy management of urban districts, learning systems can help to smooth load peaks, better integrate renewable energies and make efficient use of storage resources. For example, an algorithm could learn when and how much electricity should be temporarily stored in battery storage systems or released in order to prevent grid overloads and reduce costs. Such projects are already underway in Amsterdam, Vienna and Munich – with growing interest in medium-sized German cities too.

Another field is urban environmental management. Here, reinforcement learning can be used to optimize the irrigation cycles of green spaces, dynamically adapt air pollution control plans or even predict the spread of heat islands and initiate countermeasures. This is particularly exciting in contexts where many small measures interact and the system must constantly react to new weather conditions or usage patterns. The algorithm becomes the guardian of the microclimate – and learns more every summer.

Reinforcement learning also opens up new avenues in the area of citizen participation. The city of the future is not a black box, but an open system in which a wide variety of players exert influence. Here, learning algorithms could be used to analyze and structure citizen feedback, complaints or suggestions and feed them into decision-making processes. This will not only make the city smarter, but also more responsive – provided that the processes remain transparent and comprehensible.

Finally, it is worth taking a look at disaster prevention. Whether heavy rain, heatwaves or cyber attacks on critical infrastructure – reinforcement learning can help to identify crises at an early stage and develop adaptive emergency strategies. The algorithm simulates various scenarios, evaluates the effectiveness of measures and adapts its behavior to the situation in real time. In theory, this sounds like a magic bullet – in practice, we are still at the beginning. But the direction is clear: the more complex and dynamic the challenges, the more valuable learning systems become.

Opportunities and risks: Between smart city and black box

As promising as reinforcement learning sounds, it also harbors considerable risks. The biggest problem is traceability. Many algorithms are so complex that their decisions are barely comprehensible to outsiders. In the worst case, a black box is created in which nobody knows exactly why a certain measure was taken. This is a real dilemma, especially in the public sphere, where transparency and democratic control are high priorities. Who wants an algorithm to decide on construction site detour, closure times or resource allocation without any explanation?

Another risk lies in the quality of the data. Reinforcement learning is only as good as the information it receives. Distorted, incomplete or misinterpreted data can lead to the algorithm learning the wrong strategies – with serious consequences. A classic example is traffic models that are optimized exclusively for car traffic and systematically disadvantage cyclists or pedestrians. There is a risk of a technocratic bias here, which reinforces existing inequalities rather than reducing them.

The issue of governance should not be underestimated either. Who decides what goals a learning system pursues? Who owns the data, who controls the algorithms, who is liable in the event of damage? Many cities still lack clear rules and standards for the use of reinforcement learning. Here, local authorities, legislators and scientists are equally challenged to develop guidelines that enable innovation but also prevent misuse.

Nevertheless, the opportunities outweigh the risks if the framework conditions are right. Reinforcement learning can help to make cities more flexible, more resource-efficient and more liveable. It offers the opportunity to dynamically adapt planning processes to changing conditions and thus create a new quality of resilience. Those who are aware of the risks and actively manage them can benefit enormously from the learning city – and establish a new culture of urban innovation in the process.

Above all, it is important to ensure acceptance among the population. Without transparency, explainability and participation, any AI application in urban areas will fail. The city of the future must not only be intelligent, but also fair, open and inclusive. Reinforcement learning is a powerful tool – but it remains just that: a tool that must be used with caution, a sense of proportion and democratic control.

Technical and organizational requirements: What cities need in order to learn

For reinforcement learning to develop its potential in urban planning, it needs more than just a powerful server in the basement of the town hall. First of all, high-quality, up-to-date and diverse data sources are essential. Whether traffic counts, energy monitoring, weather data, citizen feedback or socio-economic indicators – the broader and more networked the data basis, the better the algorithm can learn and make meaningful suggestions. This is where urban data platforms and digital twins often come into play, acting as central data hubs and bringing together different sources.

But data alone is not enough. Powerful sensor technology and infrastructure that can deliver real-time data is crucial. Many German cities are still at the beginning here – keyword smart city. Comprehensive IoT networks, open interfaces and standardized data formats are the basis for any form of reinforcement learning. Without them, many projects remain stuck in the pilot stage and never develop their full impact.

Organizational embedding is at least as important. Reinforcement learning not only changes technical processes, but also administrative and planning processes. Responsibilities need to be clarified, interdisciplinary teams formed and new skills developed. Planners, IT experts, data analysts and lawyers are needed to work together on solutions and actively shape the change. This shows that the introduction of learning systems is always a change process that requires openness, courage and a willingness to experiment.

Not forgetting the legal and ethical framework conditions. Data protection, IT security, responsibility and liability must be considered from the outset. Especially in Germany, where awareness of these issues is rightly high, clear guidelines are essential. They create trust and ensure that innovation does not become an end in itself, but brings real added value for the city and its citizens.

In the end, it is governance that determines success or failure. Only if cities are prepared to take responsibility, clearly define goals and make the use of reinforcement learning transparent can the technology develop its potential. It is a challenging path – but those who take it will reinvent urban planning from the ground up.

Practical examples and outlook: How the learning city of tomorrow will work

A look ahead shows: The first cities are already successfully experimenting with reinforcement learning – including in German-speaking countries. In Vienna, for example, a pilot project for the intelligent control of street lighting is being tested. The algorithm learns when and where light is needed and dynamically adapts the lighting to the volume of traffic and environmental influences. The result: less energy consumption, greater safety, happier citizens.

In Munich, reinforcement learning is used to better understand the behaviour of pedestrians at busy intersections and to optimize traffic light phases accordingly. This shows that learning systems can bring benefits not only for drivers, but for all road users. A further step towards truly multimodal mobility planning.

There are also exciting approaches in water management. In Rotterdam, a reinforcement learning system is being used to improve the control of pumping stations and retention basins during heavy rainfall events. The aim is to prevent flooding and manage water efficiently – an issue that is also becoming increasingly important in Germany due to climate change.

Internationally, cities such as Singapore, Toronto and Barcelona are pioneers in the use of reinforcement learning for urban systems. They are using the technology to better manage traffic flows, energy consumption, environmental pollution and even social dynamics and respond to new challenges. The lesson: the courage to innovate pays off – and imitation is expressly permitted.

For German cities, now is the perfect time to take the first steps. Small pilot projects, interdisciplinary teams and an open error culture are the key to success. Those who learn from the best and gain their own experience can make targeted use of reinforcement learning to make the city of tomorrow more resilient, liveable and sustainable. The technology is ready – the only thing missing is the courage to use it.

Conclusion: Reinforcement learning – trial and error as a planning principle of the future

Reinforcement learning stands for a new generation of urban planning: adaptive, data-supported and dynamic. The learning algorithms make it possible not only to analyze urban systems, but also to actively control and continuously improve them. From traffic control and energy management to citizen participation, the method opens up undreamt-of potential – provided it is used transparently, responsibly and in a targeted manner.

Of course, the technology is not a sure-fire success. Without high-quality data, clear governance and an open culture of experimentation, there is a risk of failure due to the complexity of the city. But those who embark on this journey can learn from their mistakes – and transform urban planning in the long term. The learning city is not a distant dream, but a real opportunity for all those who are prepared to accept trial and error as a new principle. It is time to stop seeing urban reinforcement learning as a gimmick and start seeing it as a tool for the future of the city. Because the cities of tomorrow are not created on the drawing board, but in the data streams, simulations and learning effects of the present.

18.12.2024

Product

Wood – an urban material ?

Building design

Wood in the cities – there are a number of arguments in its favor. The material is CO2-neutral, has good insulating properties and is a renewable raw material. Architect and civil engineer Wolfgang Winter would design any new building out of wood. Sufficient material and the technology to build upwards are available.

Wood in the cities – there are a number of arguments in its favor. The material is CO2-neutral, has good insulating properties and is a renewable raw material. Architect and civil engineer Wolfgang Winter would design any new building out of wood. There is enough material and the technology to build upwards.

Baumeister: Mr. Winter, we are confused: on the one hand, we hear about a renaissance in timber construction, but on the other hand, timber construction in the city has declined. Which is true?
Wolfgang Winter: A stable market segment has emerged for single-family houses in Central Europe. In multi-storey construction, it is more complicated: in the 70s to 80s, i.e. after the war, there was a market share of zero. In Austria, Germany and Switzerland, state-subsidized campaigns were created at the time to accommodate the returnees from Russia – building was done with wood. These campaigns caused the market share to rise to five percent in the short term. The fact that this figure is now weakening again is due to the lack of funding. The question is: Can ecological measures that cost more than concrete construction be justified at all? This brings up the concept of affordable housing, because expensive construction is not socially sustainable. Then we just build in concrete again. From this perspective, social sustainability excludes ecological sustainability.

B: Does timber construction necessarily have to be more expensive?
W W: In the short term, yes. A cubic meter of concrete costs 50 euros. Wood, on the other hand, costs 400 euros per cubic meter. So if you replace concrete with wood in an equivalent construction project, it is more expensive. That is of course a disadvantage of wood.

B: Where does this big price difference come from?
W W: A cubic meter of tree, as it comes from the forest, costs 100 euros. The price is determined by the forester who cuts the wood and the forest owner who waits 100 years for the tree to grow. If the tree is sawn down, 50 percent is lost through the waste products. This means that a cubic meter costs 200 euros. The wood then has to be dried and glued, tempered and quality sorted. This is always a high cost for a natural product.

B: The solution?
W W: You have to build intelligently. For timber construction in the city, you need a well thought-out system and a quality-assured product. This is not possible in this DIY niche with a regional, “cute” timber construction culture. For large-scale industrial projects with 200 residential units that need to be completed within six months, you need prefabricated products. In terms of price, timber is competing with in-situ concrete poured on site. At the moment, it is still losing this battle.

B: So timber has a lot of competition. Until 1800, things were different – every building was made of wood, at least in part. When exactly did the turning point come?
W W: Until 1800, all construction was “self-build”. People built with the materials that were available on site. Carpenters and bricklayers built without architects. A big break came with industrialization. The crafts disappeared. The railroad, steel and cement arrived.

B: What’s more, in the 19th century there was simply no more wood…
W W: That’s when the laws for sustainable forestry were introduced. From the second half of the 19th century, they stipulated that if a tree was felled, two new ones had to be planted.

B: So we would have enough wood again today. And the “paperless office” will surely ensure even more wood…
W W: The paper thing is not so easy to conclude. In fact, the yields from forests have increased enormously. This is due to properly managed forests. Until the 18th century, yields were five cubic meters per hectare. With forest management, the figure climbed to 10-15 cubic meters per hectare. Due to climate change and the high CO2 content in the air, forests are becoming even more productive.

B: So we would have enough wood to theoretically build entire cities with?
W W: Yes. There is more wood growing than we need. If we wanted to, we could build every new construction project in wood.

B: How high could we build with wood?
W W: Wood has a compressive strength of 30-40 newtons, concrete also has 30 newtons. Of course, it has a lower tensile strength than steel. But this can be compensated for with a higher cross-section. And timber is still relatively light. Pure timber buildings of up to ten storeys are technically possible without any problems, even when fire protection requirements are taken into account. Fire protection is actually a question of escape routes and access and not the combustible material.

B: Especially when we’re talking about urban areas, isn’t there a great risk of fire spreading from one building to another?
W W: Every fire is started by mobile fire loads – the furniture, the curtains. Wooden buildings don’t burn any more than other buildings. Wood does not ignite more quickly, nor is the risk of a fire starting greater than with other building materials. The most important fire protection measure is the escape routes.

B: Timber construction seems to reach its limits at ten storeys. Why then want to build even higher? Shouldn’t we think about the material according to its use?
W W: The tensile forces are the problem. But you can use timber steel for that.

B: Wooden steel?
W W: When we talk about timber-steel construction – steel clad with wood – then it’s the same principle as with reinforced concrete: you have a large cross-section consisting of compression elements, in this case made of wood, and inserted flat bars or angles that absorb the tension. From a structural point of view, all skeleton structures that are currently made of reinforced concrete could be made of wood.

B: What are the biggest advantages of timber in the city?
W W: Wood is an excellent raw material that can be used to make various products. It is easy to process. It also has low thermal expansion due to its high porosity. With other materials, you have to leave more space during installation, or the adhesive has to compensate for the expansion. Wood also has good thermal insulation properties. The advantages in the city lie in building gaps and extensions. The material is light and can be lifted into urban structures by crane.

B: Another major advantage of timber in the city is the high degree of prefabrication. Does this impose restrictions on the design?
W W: I think you can design very freely with wood. Nowadays, wood is machined and glued together. Robots mill out holes and join the wood together. So you can produce parts industrially and individually.

B: No disadvantages?
W W: Of course, it’s clear that if an architect builds monolithically beforehand, this allows for different building forms and requires different thought structures than if you put together an additive system from rods. Prefabricated timber construction requires a certain level of awareness on the part of the architect. If the architect has this knowledge, however, there is certainly freedom of design. The prefabrication of timber and steel is equivalent in the construction process. But wood has a few additional advantages.

B: Sustainability, for example. However, the word is now used everywhere. Has it lost any of its strength as an argument for timber construction as a result?
W W: A lot has been smuggled into the term sustainability: architectural quality, beauty and ecology. Now we no longer talk about sustainability, we talk about resource efficiency. Timber construction itself is clearly resource-efficient. And since we change our building fabric in relatively short cycles, resource efficiency also means what the material makes possible in terms of later use. The monolithic cast construction cannot be dismantled and rebuilt elsewhere. Steel and wood are easier to recycle.

B: Do you think that in a world surrounded by technology, we are longing for a natural building material?
W W: Yes, that is certainly part of it. On the one hand, there is this useful timber construction, but it doesn’t claim to be a statement. Our urban buildings have many half-timbered structures that were subsequently clad. Today, of course, things are different. Since concrete was the building material of the 20th century, if you offer an alternative, you also have to work with a feeling: We now live in a material that is closer to nature. But that will certainly only remain a niche. Eco-awareness is a decisive factor for a maximum of 20 percent of the population. The others don’t care if they live in a concrete building.

B: You said that concrete was the dominant building material of the 20th century. Is wood the building material of the 21st century?
W W: Wood has everything it takes to become the building material of the 21st century. Concrete was the building material of the 20th century, especially in Europe. This has to do with our specific history, with the Second World War. You could argue that the population’s growing environmental awareness is the basis for wood becoming the material of the 21st century. But, of course, you have to see how strongly wood is being fought over by the forestry, paper and pellet industries. The competing players for this natural material must agree that it makes the most sense to build with wood.

Searching for clues on Slate Islands

Building design

The poetry collection "Schiefern" by Esther Kinsky explores the analogy between human memory and metamorphic rock. Photo: Suhrkamp — The poetry collection "Schiefern"

The poetry collection “Schiefern” by Esther Kinsky explores the analogy between human memory and metamorphic rock – a sensual search for the lifeless. On the map, they are small patches off the west coast of Scotland, so small that it is easy to overlook them. You have to seek them out specifically to find them. You don’t just come across […]

The poetry collection “Schiefern” by Esther Kinsky explores the analogy between human memory and metamorphic rock – a sensual search for the lifeless.

On the map, they are small spots off the west coast of Scotland, so small that it is easy to overlook them. You have to seek them out to find them. You don’t just stumble across them. The Inner Hebrides of Scotland, a group of islands at the top of the British Isles, are a popular travel destination. Those who come here long for the original, the wild, the rugged. For the salty wind that catches hair and clothes and makes them stiff. For the Atlantic, its waves crashing against the black rock. Gneiss. Granite. Basalt. Slate.

Esther Kinsky, translator and poet and 2018 for “Hain. Geländeroman” in the fiction category at the Leipzig Book Fair, has dedicated a volume of poetry to slate and the region where the sedimentary rock was mined for centuries with the simple yet telling title “Schiefern”.

The quarries on Slate Islands are still there, as are the remnants of a now defunct industry. Kinsky embarks on a voyage of discovery and wraps her observations of nature in words that are enigmatic to decipher and carry us away to the remoteness of the Inner Hebrides, to the black, raging sea, above which the reader floats like an invisible person in the mental space that Kinsky spins with her words.

It is precisely there, in this space of thought, that the analogies between something thoroughly lifeless and human can be found. There are only a few people in this three-part volume, but it is not lacking in humanity. In fact, it is quite astonishing how sensually it is possible to write about waves carrying spray and “plates with a / surface like petrified quiet waves” without slipping into kitschy romanticism.

“Nature Writing”

Nature has been tempting writers to write about it as the main protagonist since the 18th century. In Anglo-Saxon, “nature writing” is the name given to lavish literary descriptions of trees, meadows, flowers and cloudbursts. In German, the term “Naturpoesie” or “nature poetry” has become commonplace. Esther Kinsky has stood out in literature for years with such nature poetry.

In 2013, she weaved four cycles of poems about decay and growth in “Naturschutzgebiet” (Nature Reserve), based on a neglected city park. If Kinsky’s work is now categorized as “nature writing”, she is happy to contradict this. In an interview with Deutschlandfunk radio, she once said that she did not see herself in the tradition of nature writing. This term is too diffuse, too sprawling in terms of what it encompasses and what it does not. “Nature writing” can be anything, she says. So why not her latest work “Schiefern”, one might ask?

The layers of time

Early on in “Schiefern”, the word “memory” is used “as a space of absences, moved by the transparent hand of unpredictable synapses and imponderable shifts of deposits in the slowly emerging and deepening furrows and folds of the brain”. Kinsky is concerned with the layers of time that accumulate over memories. At first very gently, then more clearly, she draws linguistic parallels between human memory and the preserved history on the surface of the rocks, which the tides and times have passed by over millions of years.

The past is preserved in the stone, it only has to be read from its wrinkles, as if the stone were an old, cherished old man whose weathered face bears the traces of life. Kinsky writes of “signs without hand or foot / in the stone to which no one / knows how to make a rhyme / but the greatest possible past”.

“Schiefern” could be the modern sequel to Adalbert Stifter’s 1853 short story “Bunte Steine” and join the ranks of “Granit”, “Kalkstein” and “Turmalin”. But as treacherously idyllic as Stifter’s detailed, Biedermeier-like depictions of nature are, Kinsky’s description of the Slate Islands is just as uncharitable. The coolness of the surroundings snows through her words. There is a harshness in them that you don’t want to imagine without.

Information about the book

Esther Kinsky: Slates.
D: 24,00 Euro
A: 24,70 Euro
CH: 34.50 Swiss francs
Published: 23.03.2020
Hardcover, 103 pages
ISBN: 978-3-518-42921-1