Successful urban planning needs to satisfy a complex combination of aesthetics, practicality, and human needs.
A recent study found that AI urban planning tools are better and much faster than human city planners at a number of the more mundane aspects of urban planning.
In 2020, Carlos Moreno introduced the idea of the 15-minute city, where residents are never more than a 15-minute walk or bike ride from all the services they need. The simplicity of the idea belies the complex planning it takes to achieve these efficiencies.
Automation scientist Yu Zheng and his colleagues from Tsinghua University in Beijing wanted to see if machine learning could be an effective urban planning tool.
They developed an AI system that could optimize land use and road layouts to align with the principles of the 15-minute city while satisfying local planning policies and needs.
They started small by having the AI design an area of around 3×3 city blocks, or a few square kilometers.
The results of their experiments showed that when it came to access to services, green spaces, and traffic levels, the AI was able to outperform human designs by about 50%.
And the AI was a lot faster at these tedious iterative optimizations. It was able to complete tasks in seconds that normally took human designers about 50 to 100 minutes.
Commenting on the study, Paolo Santi from MIT said that urban planning is “not merely an allocation of space to buildings, parks and functions, but the design of a place where urban communities will live, work, interact and, hopefully, thrive for a very long time.”
Zheng and his team agree with those sentiments and see AI as an assistant to urban planners, rather than a replacement.
When they compared human-only designs to human-AI ones, they found that using AI collaboratively increased access to basic services and parks by 12% and 5% respectively.
The researchers surveyed 100 urban designers and asked them to evaluate a number of designs without knowing which had been designed by AI or by humans. In some cases, there were no clear preferences but some of the AI spatial designs received substantially more votes.
The AI model required 2 days of training to manage the 3×3 block design. Increasing the space to a 4×4 block design requires twice the amount of planning decisions. Scaling an AI tool like this one to actually plan an entire city effectively is some way off, but the potential is obvious.
If urban planners use AI to do the tedious computational optimizations it will give them more time to pay attention to aesthetics and societal needs that are more difficult to quantify when training an AI model.
