So far, we have introduced the topic of AI into Architecture. I have also claimed that our building code may be governed by sub-optimal learning structures. In this room we will therefore explore the activation of our past, to see if we can extract old memories to improve future architectural learning.
CHAPTER 3 OVERVIEW
Chapter 3: Exploring built collective memories
“Those who cannot remember the past are condemned to repeat it”
I believe an important step to improve the overall architectural quality in our built environment, is to facilitate for better understanding of how the changes we make to the building code take effect in our buildings. Can we utilise artificial learning to better understand this complex dynamic? Could our old projects be memories of experience for an improved learning process? In this chapter we will research how the current situation affords us the opportunity to, or currently hinders the advancement of, leveraging our experience and our past. We will explore this with the hopes of later experimenting with an AI’s ability to grasp architectural representations. Therefore, we will begin by choosing a model for machine learning.
3.1 A Machine Learning Algorithm
Having clearly set a direction for our exploration, we will later elaborate upon our needed constraints. The choice to focus on plan layouts will be an important decision. To see if we can in fact train a machine learning algorithm to understand architectural representations, we will also need data (the technical term for what will be our architectural layouts) and a suitable algorithm to grasp the underlying patterns.
For our experiment, we will use a generative machine learning model called “pix2pix”. It was first developed and researched by the AI lab at Berkeley in 2018.1 This work was built on the seminal research by Ian J. Goodfellow et al. in 20142, and the work completed later that year by Mehdi Mirza on the same topic.3 The pix2pix utilises what is called conditional adversarial networks in order to generate novel pictures, from the patterns in the training data. A high definition implementation of pix2pix is available (affording the user much higher picture resolution opportunities), but since we are looking for proof of concept, rather than some finished product, we will stay with the first version for simplicity’s sake. Consider this the briefest of introductions, as we will go into the details of the specific algorithm, and the mechanisms of its functioning, in the next room.
In support of our specific choice of machine learning model, we may look to work done at ETH Zurich4, a recent master thesis from Harvard5, and research from the University of Pennsylvania and Tsinghua University in China.6
1. Phillip Isola et al. (2018), “Image-to-Image Translation with Conditional Adversarial Networks”, Berkeley AI Research Lab
2. Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozairy, Aaron Courville, Yoshua Bengioz (2014), “Generative Adversarial Nets”, University of Montreal
3. Mehdi Mirza (2014), “Conditional Generative Adversarial Nets”, University of Montreal
4. Morteza Rahbar, Mohammadjavad Mahdavinejad, Mohammadreza Bemanian, Amir Hossein Davaie Markazi & Ludger Hovestadt (2019), “Generating Synthetic Space Allocation Probability Layouts Based on Trained Conditional-GANs”, Applied Artificial Intelligence, DOI: 10.1080/08839514.2019.1592919
5. Stanislas Chaillou (2019), “AI+Architecture: Towards a New Approach”, Harvard Graduate School of Design
6. Weixin Huang, Hao Zheng (2018), “Architectural Drawings Recognition and Generation through Machine Learning”, Tsinghua University, University of Pennsylvania
3.2 Empirical Consequences of Changes to Law
To better understand the consequences of new updates to the building code, our past is the only option of study. As luck would have it, it happens to be that we have, on the topic of the building code, a large set of questions regarding consequences and effects; of which all contain their directly applicable and connected answers. Indeed, we could ask what the consequences were on architecture and the built environment, from every single change and update, and from any specific previous period in law. We could, for example, question: “What was the built effects of the active PBL and the TEK10, from 2010 to 2015?”. An answer then lies directly encapsulated as all the projects, given their context, completed within this period.
Following this, we quickly realise that we have a vast set of laws and accompanying technical regulations, all connected to their separate time periods, of which have their own sets of built consequences. In other words, we have two separate and connected structures of information. One of symbols, represented by text and numbers, in the building code, and one of geometric physical structures outside our windows. These are clearly connected through cause and effect. However, given the complexity and sheer volume of information, we have for quite some time, in a practical and computational sense, been shut off from studying their overall connected patterns. Today, such complex structures of information are, if properly prepared, exactly what AI offers to help us with.
3.3 Applying Necessary Constraints
To strategically limit the scope of our investigation, let us choose some constraints for the continued research. As previously mentioned, we will narrow our explorations of the architectural representation to the architectural plan layout. These drawings or diagrams are at the heart of the architectural practise. They clearly explain, with accurate measurement, the sequences, relations, and proportions of built construction and resulting composition of space.
For precision, we will also limit our search with respects to typology. The building code enforces constraints to all types of buildings, but we will focus on housing for now. Statistics for April 2020, by Statistics Norway, informs us that there are 2,610,040 residences in Norway. Of these, there are 1,281,004 houses, 311,648 rowhouses, 235,467 twin-housings, and 643, 631 apartments.7 Perhaps of little surprise, we see how single-family houses are highly favoured amongst Norwegians. However, we shall later attempt artificial architectural learning on our collected plan layouts. Therefore, residences with one floor only may be preferable as a starting point, to avoid unnecessary complexity.
Also, single-family houses, in their current form, may prove holistically unsustainable for a future in need of more efficient space and material use. Apartments often offer greater space efficiency, both in terms of their inner functioning, and as a part of other stacked residences. They are also more common to urban settings. These areas are marked by a higher living density, often being more in line with sustainable living solutions for the future. We also know that urban growth is likely to continue. Let us therefore chose to investigate apartments and specify it to “3-room apartments”. This versatile type can work as a mean-type for exploration. Being a relevant residence to a large target audience in many stages of life.
Exploring all building code changes would also be impossible with the current time available. I deliberated in discussions with Trondheim municipality, Standard Norway, and professors at NTNU, amongst others, in order to choose which periods might be of most interest. Resultingly, I have chosen 3 time periods for us to study. Each characteristically marked by their regulations on technical requirements. Since we are seeking to read the built changes through time – through architectural plan layouts – the choice of using technical regulations as markers is plausibly a reasonable one for this first experiment. The following periods were chosen:
• 1. 1987-1997 (TEK87)
• 2. 1997-2008 (TEK 97)
• 3. 2008-Present (TEK10, TEK17)
These three cohorts consist of varying building code and separate technical regulations.8 Therefore, providing us with interesting variations, given their relational evolution. They are also relatively new cohorts. From this, we shall assume that finding material will be less time consuming than extraction of plans from older cohorts.
Another reason for this choice was providing a view to the interesting changes related to universal design. A concept that entails the planning of products, environments, programs and services, in a way that invites for use by as many people as possible, in an fair and equitable way (specifically relevant for architecture, with regards to those with impairments related to movement).9 Also, the evolution through our cohorts provides interesting change with respect to elevator requirements, energy demands, sprinkler-systems, daylight requirements, fire security, chimneys, apartment storage, and much more. Most of which we might track, as they would influence the plan layout.
From previous research done with the pix2pix-model, an absolute minimum of 300 apartment layouts were used in the study from ETH Zurich. However, aiming at 700 or more plans for each cohort has a track record of success, with the specific method that we will be using.5 It is therefore our target moving forward.
Since we want to explore the actual built consequences, we will importantly only gather “as built” architectural plan layouts. We will also effectively limit the plan extraction, choosing only the apartments built two years into, or later, within their respective cohorts. In doing so, we may increase the certainty that the rules have had time to take effect. As an example, we will only choose apartments from 1989-1997, when we are gathering for the 1987-1997 TEK87-database.
7. Statistics Norway (SSB) (2020), Accessed: April 2020, Available at: https://www.ssb.no/boligstat
8. Note: Complete building code history, only available in Norwegian, accessible from: https://www.regjeringen.no/no/tema/plan-bygg-og-eiendom/bygningsregelverket-fra-1965--20172/id2590706/?expand=factbox2598937
9. Inger Marie Lid (2020), «Universell utforming», Store Norske Leksikon, Available at: https://snl.no/universell_utforming
3.4 Initial Search: A First Discovery
When beginning the search for architectural layouts, I sketched a few different paths. As a first step I reached out to most of the major housing contractors. Parallel to this, I went through just under 900 relevant marketing advertisements, for appropriate apartments on www.finn.no. Finn.no is the current leading sales- and advertisement platform for real estate in Norway.10 After building a library of the most viable 150, out of the 900 possible layouts found, I made the decision that no apartment layouts contained in these advertisements were of sufficient quality for our continued research. Specifically so, with respect to details, representation, and trustworthiness, such that we could warrant them for our use. Nonetheless, the work gave insight and ideas into the topic of standardisation.
An important discovery was also made early, upon contacting and discussing the layout search with a large set of the top 10 housing contractors in Norway. Seemingly, none of the companies spoken to, had organized their previous projects in such a way that they could efficiently and properly extract their potential value. Additionally, an overwhelming amount of the older projects were still not digitised. Much of the content was available through difficult means of extraction, either in separate project folders, across different servers or as combinations of digital and analogue versions.
Seemingly, even the large and leading organisations within the building industry had yet to fully understand the potential value of their past projects. Therefore, our sought-after material did not yet have the structure and organisation needed for an efficient extraction. It was neither easily prepared for raw extraction, nor in any way prepared for predictive data analysis and machine learning. The latter is potentially a question of lacking work towards ai-standardisation. An important discovery that requires urgent attention and will be addressed further in chapter 4.
10. Finn.no, owned by the e-commerce company Schibsted.
3.5 Building a Database: Journey to the Archives
After unsuccessful attempts at extracting enough architectural layouts, either digitally across Norway or locally in Trondheim, contact was established in February with OBOS Oslo. OBOS is Norway’s largest building- developer and maintenance organisation. It is owned by its members, and in 2016 it held over 400 000 members, while managing over 215 000 residences.11 In early march, I met up at Ulvenveien 82 in Oslo, and was given full access to their large archive. It covered a full storage hall, containing projects from all the way back to their beginnings around 1929.
OBOS Oslo had already converted parts of their archives into digital copies. They had also received offers for a complete digital transformation, but seeing how the value of these old projects were somewhat unclear, the perceived high price had already postponed full initiation. For many projects within our chosen cohort, digital copies were, as luck would have it, already in place. This meant less time scanning documents than first anticipated.
Extracting the full range of the built memories within the archive walls, is at first sight, presumably as time consuming as it is possibly important. Having seen the current archive-state of a professional actor, such as OBOS Oslo, gives a good sense of the initiative and effort needed to activate the majority of past projects on a national basis. What became clear, was also the need to combine the future extraction of such archives, with the standardisation process needed for AI- and data prediction analysis.
11. Norske Boligbyggelags Landsforbund, NBBL, Report 2016, 2016, Accessed: May 2020, Available at: https://www.nbbl.no/media/11466/2016-aarsstatistikk.pdf
3.6 Extraction Results
After around 10 days at OBOS, the relevant architectural layouts were extracted. The result was just over 500 3-room-apartment layouts. These were accumulated from 170 different individual housing companies, of which OBOS administrates. Companies were initiated from 1986-2019, and layouts can be assumed to represent “as-built” consequences of their building-code period.
• 1. 1987-1997 (TEK87): 135 Architectural Layouts
• 2. 1997-2008 (TEK 97): 184 Architectural Layouts
• 3. 2008-2020 (TEK10, TEK17): 183 Architectural Layouts
With respects to wishes from OBOS, the list of collected housing companies, and their individual addresses, will be kept off record. Most of the apartments were located in Oslo, or close to the Oslo/ Viken region.
In the start of this chapter we initially aimed for around 300-700 plans for each cohort. As such we are clearly off the mark. Going through the entire relevant OBOS database, we would perhaps have assumed to find more layouts, however, with strict requirements to quality, type, and uniqueness, 502 plans are for now what we have. Having initially aimed for far greater collection of plans for each cohort, we will likely have to adjust our research going forward in the next chapter.
• We have chosen the pix2pix-model, a generative machine learning algorithm, to further experiment with architectural learning.
• Old projects could serve as valuable memories. As such, containing lessons for learning the cause-effect relationships between building code updates and the resulting consequences on building practises and buildings themselves.
• We have applied necessary constraints to our experimental architectural learning research. For this project, we will focus on gathering architectural plan layouts of three-room apartments, from three specific building code time periods.
• Upon contacting relevant agents in housing construction, we saw how the preparatory work to extract past projects, are not meaningfully set in motion. The need for an official AEC AI-Standard, to begin building future databases, may be critical.
• Around 2 weeks were spent extracting plans from the archives of OBOS Oslo. The result was just over 500 plans in total, from the 3 different building code periods. This is far less than we presumably need. Going forward, we will therefore necessarily adjust our explorations.
“There is no such thing as a failed experiment, only experiments with unexpected outcomes.”