Self-Organizing Floor Plans

This article introduces and comments on some of the techniques currently used by designers to generate automatic building floor plans and spatial configurations in general, with emphasis on machine learning and neural networks models. This is a relatively new tendency in computational design that reflects a growing interest in advanced generative and optimization models by architects and building engineers. The first part of this work contextualizes self-organizing floor plans in architecture and computational design, highlighting their importance and potential for designers as well as software developers. The central part discusses some of the most common techniques with concrete examples, including Neuro Evolution of Augmenting Topologies (NEAT) and Generative Adversarial Networks (GAN). The final section of the article provides some general comments considering pitfalls and possible future developments, as well as speculating on the future of this trend.


Media Summary
This article introduces some of the techniques currently used by designers to generate automatic building floor plans and spatial configurations in general, by using machine learning and neural networks. This is a relatively new tendency in computational design that reflects a growing interest in advanced generative and optimization models by architects and building engineers.
In this work we analyze some of the most popular techniques to automatically generate spatial configurations in architecture. We try to highlight how and why different techniques are used and provide comments on the results for each approach.
We contextualize these methods within previous work of self-organizing configurations and suggest how we should consider and use these techniques in the future.

What Are Self-Organizing Floor Plans?
This article discusses how the space in which we live can be designed by algorithms instead of humans, with designers working out their projects driven by computer logic instead of Euclidean geometry. Of course, the question is more elaborate that it seems and it is, probably, even less exciting.
In the context of design, floor plan is an architectural term that indicates the bidimensional spatial arrangement used by designers to determine internal building layouts. More generally, people refer to floor plans today as blueprints, spatial layouts, internal spatial arrangements, and so on.
Usually, designers start the spatial organization of space (a room, a building, an entire city) by encapsulating the main spatial qualities in an initial sketch. Through several procedural stages, this sketch is then developed into spatial arrangements, floor plans, and technical drawings (Plowright, 2014). During this nonlinear process, the designer evaluates each development of the drawings against the prescriptive data (building regulations, health and safety, minimum distances, etc.) and subjective aspects (taste, experience from previous projects, and experimental concerns).
Although automation in architectural and design is an integral part of the history of design (think of Sigfried Giedion's [1948] comprehensive accounts of how "Mechanization takes Command"), traditional approaches where projects start with sketches and templates and are developed through continuous refinements, are today challenged by a more linear and data-driven workflow where designers use algorithms to produce their project (cf. Ferreira & Leitão, 2015).
Unlike traditional architects, computational designers need to start their work by modeling the design problem (what is required by the project brief), including all relevant information (from building regulations to dimensional guidelines) and elaborating a logic that allows all parts of the design to be hierarchically related and processed. Once the design logic has been elaborated, this approach requires the use (or the development) of an algorithm of which implementation will allow the final design to be computed. A direct application of this method is represented by selforganizing floor plans (SOFP).
In the context of this article, with SOFP we refer to the combination of methods that designers use to automatically generate spatial building layouts. Generally, these methods include machine learning (ML) and optimization techniques, neural networks (NN) and evolutionary algorithms. The models usually include topological approaches (for example, with Kohonen's self-organizing feature maps [SOMs]) and object recognition (e.g., deep neural networks). The notion of self-organization has different meanings depending on its context. In the study of complex systems, for example, selforganizing may be associated with the idea of emergence, where a global order among elements is achieved through interactions of the parts governed by local rules (think of fractals or cellular automata). Although some of the topological approaches (for example, Kohonen's SOMs) rely on this idea, the examples included in this article also include different approaches, where self-organization is achieved by using neural networks and genetic algorithms.
To understand the importance and relevance of SOFPs for architecture and design today, we need to contextualize them within the design industry at large. In fact, there are three main viewpoints that can help to create a comprehensive picture of what is usually called a 'self-organizing floor plan.' Firstly, the media (including magazine targeting nonarchitects or data scientists) presents SOFPs as a magic tool where designers input generic data into a black-box algorithm that provides fully fledged floor plans as an outcome. This has led to a large debate (which is still ongoing at the time of writing) about the future of designers and the actual need for them in a situation where computers can potentially substitute them entirely (for example, Carta, 2020a;Celento, 2007;Fairs, 2019).
To date, we are still far from what the media often portrays where computers automatically produce floor plans and building designs. We will probably never arrive at this point as design is an intrinsic human activity that requires intuition, taste, and, ultimately, a good degree of subjectivity (cf. Davis, 2020). However, it is also true that synthetic approaches can be very helpful to designers as they increase the accuracy of their work, reduce the number of errors in a project, save precious time, and avoid tedious tasks. This opens up new possibilities for development (Carta, 2020b).
Opinions differ among architects and designers on this matter, between those who recognize the potential of the use of algorithms in the design process, and those who consider them with suspicion. The relationship between human and machine in architecture is elegantly summarized by David Rutten as: "the architect remains the designer, while the computer becomes the critic" and this is related to the idea that "ultimately, a computer has no idea what it's doing, let alone why it's doing it.
Understanding stuff is what humans are good at" (Rutten, 2021).
Rutten's opinion indirectly suggests the importance of considering hybrid approaches, where the contributions of machines and humans are deployed at different stages and to tackle different problems in a design process, playing on each specific strength. A growing number of projects epitomize this human-AI (or human-machine) co-creation (e.g., Pitso, 2019;Wikström, 2018;Woo, 2020).
The examples included in this article illustrate this delicate relationship between designers and computers, which is far from being fully resolved. One way of looking at this is through the lens of the designer who sources and selects the initial data sets, curates and cleans the data in preparation for the computer to be calculated, and models the design problem in order to address the questions at stake (from the design brief, the client's requests, etc.). In this case, we may argue that the designer is still in utter control of the design while the machine is merely executing instructions set by the designer. Another perspective can be from the viewpoint of the computer that is employed in the design process in the hope that new and unforeseen results may emerge. Take the example of clustering a given data set. Designers may use clustering or classifying techniques to find new aspects in their design that would not be visible without the aid of a computer. In this case, we may argue that the designer is still in control of the design, although they trust the computer to generate parts of the design that are outside of their reach. Designers become meta-designers or, in extreme cases, they become critics of their own work, as suggested by Rutten.
The argument suggested by this article is that designers in both cases are still in full control of their creative work, especially considering the effort they make in preparing and curating the data set, modeling the problem, setting up the entire workflow, evaluating results at any stage of the process to ensure consistency and validity and, ultimately, being responsible for the final results obtained.
Secondly, designers, especially those interested in computational design (that is, simplifying, a subset of design where computers are heavily used), consider selforganizing plans and generative design in general as one of the future directions for development. Not only do computational designers create their own tools using openended graphical programming interfaces but they also see such tools as increasingly integrated into the main software packages that are largely used in the design industry (for example Autodesk Revit or McNeel's Rhinoceros 3D). Designers are moving their work (together with their creativity and problem-solving approaches) from the traditional representation of initial ideas and their continuous refinement to a radically new dimension. In this, designers create their own (computational) tools that serve them as intelligent collaborators in their projects (Jabi, 2013;Terzidis, 2006).
Thirdly, computer scientists have found that applications of certain models and algorithms can be tested against a real-case scenario (for example, the design and construction of a building). An example of this is the work that Tarabishy and his colleagues have done on the use of "deep learning surrogate models for spatial and visual connectivity" (Tarabishy et al., 2020). In this, they use a number of ML methods to reduce the computational time needed to simulate the spatial and visual connectivity of a given office space. More and more, we see the emergence of new consultancy practices where mathematicians, statisticians, software developers, and computer scientists work along with designers and engineers to offer bespoke and intelligent solutions for architecture and the construction industry. Examples of such practices are Kreo, Hypar, or TestFit among many others.
This article is an attempt at clarifying the reasons why the second and third views are important, for they are significantly changing some aspects of the architecture and construction industry. Indirectly, with the analysis included in this work we want to offer a robust counterargument to the first view, explaining the extent to which the idea of computers as a substitute for designers is not only simplistic, but fundamentally flawed.
In general terms, we might say that the success of each model for this particular task is connected to two main aspects. The first one is the advancements in machine learning research. The growing use of artificial neural networks (ANNs) in computational design is a reflection of the fast advancement in research in generative models (Dhondse, 2020) and true distributions (Alcin, 2019). This is as well as the increasing computational power available and training data sets available (Hodas & Stinis, 2018). As the possible applications of these methods increases and they become more reliable, research in this direction is increasingly active.
The second is the creative use that designers make of this new research. It is important to note that although many of the current projects are the result of multidisciplinary teams (including computer scientists, data scientists, and designers), they often originate from and are led by architects (see, for example, the work conducted at MIT's SenseAble City Lab). Architects, and designers in general, use computational methods to generally produce applied research using primary new knowledge produced in statistics, mathematics, computer science, data science, and so on. It is in its possible application to buildings, cities, or as in this case, floor plans that designers can offer novel and creative approaches to real-life problems.
The information-processing model that is commonly used is an ANN. One of the reasons behind this choice is that this model is proven to be particularly powerful when designers have large data sets available for their projects (for example, the website HousePlans.com offers around 40,000 floor plans that can be used as a training data set). The choice of using an NN is also quite popular due to the fact that most of the computation needed for predicting spatial distribution (floor plans) is made largely possible on standard machines (e.g., home computers, workstations, etc.) through backpropagation and general-purpose Graphics Processing Units (GPUs) (Danka, 2020).
The examples presented in this section use optimization techniques (neuro evolution of augmenting topologies-NEAT) and variants of neural networks based on graphs (e.g., graph convolutional networks-GCN). Algorithms based on graph theory result in quite effective manipulation of data in spatial configurations. Designers and those trained in spatial abstraction usually find topological approaches intuitive, for they have direct applications to spatial organization. A clear example of this is space syntax (Hillier & Hanson, 1989)-including one of its methods, isovist (Benedikt, 1979;Turner et al., 2001)-which is among the most developed and used computational spatial theories applied to architecture and cities. Modeling design problems through graph theory is a common strategy in the projects included in this section. The projects that follow include models based on graph optimization and generative models.
We start with a non-data-driven approach based on NEAT, which is a powerful optimization method. Generalizing, optimization models are employed to find solutions through an objective (or fitness to criterion) function. Possible solutions are iteratively evaluated against the objective function in search of optimal values (Hoos & Stützle, 2004, p. 13).
This approach is based on Stanley and Miikkulainen's (2002) work where an advanced method for evolving neural network topologies along with the associated weights is presented. NEAT's outperformance regarding the other fixed-topology approaches is related to "employing a principled method of crossover of different topologies, protecting structural innovation using speciation, and incrementally growing from minimal structure" (Stanley & Miikkulainen, 2002, p. 99). Simon (2017) and then Carta et al. (2020) developed a workflow to optimize an existing floor plan based on a number of design criteria. NEAT involves the use of genetic algorithms to find the most fitting topology in a given initial configuration. In this method, each solution (also called genome) is represented by a series of nodes and connections. Each node contains information that is iteratively evaluated against a fitness function to ascertain how close (or far) that particular solution is to the desired result (as the combination of the optimization criteria). Each connection stores information about the nodes that it links and their relative weights. The weights give important details about the role of that connection within the system, for example, the distance between the nodes (or rooms) or whether that link is considered within a particular instance. In each iteration, there is a mutation of the genome underpinned by the combination of the previous best-performing solutions where new nodes and connections can be added or removed. The method used by Carta et al. (2020) includes the initial generation of a random population of solutions to which a number of mutations is applied. In each iteration, the solution is mapped against a generic floor plan (that needs optimizing).
An ant-colony optimization (ACO) algorithm is used to evaluate the best connections between the rooms (corridors). The configuration is then evaluated against the fitness fiction and the criteria used as the input. The possible solutions are then used as inputs for the next generation. The process repeats using the best solutions of each generation to improve the overall solution, until satisfactory results are obtained. An example is show in Figure 1. Although NEAT applied to self-organizing floor plans can yield promising results, this method is indeed not very much used by architects and designers who seem to prefer in general data-driven approaches with NNs and generative methods.
A more common approach is graph convolutional networks (GCNs). This is where the NN informational model is based on graphs and their elaboration. They are defined as convolutional due to some of their characteristics (for example, the filter parameters), which are common to different parts of the graph (Kipf, 2016). A relevant application of GCN to floor plans is the work in progress of Li et al. (2020) where the user can input new nodes into the graph directly via a browser and the program generates a floor plan corresponding to the graph. outputs a new spatial configuration, which is optimized with regard to the initial input.
In this particular application, the input is represented by a number of points in space (position of rooms) that is read by the NN as an input graph.
The example of GCNs clearly shows how floor plans can be modeled as graphs of which topology is optimized resulting in a new spatial configuration. However, the approaches to SOFP that seem to be currently more popular are those based on generative models.
In this group we have methods where deep learning approaches are used to predict the implicit distribution and configuration of graphs (Zhou et al., 2018, p. 18). Unlike in discriminatory approaches, where features are mapped to labels, in generative models the aim is to find (by prediction) new features given an initial set of labels.
NEAT and GCN seem so far to have been successful at illustrating how initial inputs can be translated into spatial arrangements. However, it is with generative adversarial neural networks (or GANs) (Goodfellow et al., 2014) where most of the progress to date is being made by architects and computational designers. In order to understand why GANs are considered to be powerful enough for the task of generating realistic floor plans, we need to briefly see how they work.
GANs are unsupervised networks where there are two main components. The first network (generator) generates the data outputs (for example, images) that are assessed by the second network (discriminator). The discriminator is trained with both real and fabricated images to ascertain whether the images are real (true data) or fake. The generator is trained to produce images that look real for the discriminator.
The data can be backpropagated through both the generator and the discriminator to ensure that the former adjusts its parameters in order to reduce the difference between real and fake images for the latter (Goodfellow et al., 2014). States and Europe are collected from Google Street View, and cleaned and categorized into organized sets. In each training set, a pair of images is created where a raster RGB image is associated to its relative greyscale version that represents its depthmap ( Figure 5a).
The images are then formatted to be readable by the training models, StyleGAN (Karras, 2018) and Pix2Pix (Isola et al., 2017). Lastly, the two models are used to generate new synthetic images, with each of them producing different results with different interfaces.  Figure 5. Synthetic images produced with Pix2Pix (5b and 5c) and StyleGAN (5d). The volume on the right-end side of the depthmap image (5a) "is interpreted by the Rotterdam model (5b) as a large brick housing block, as is typical in the Dutch city, while the Pickwick Park model (5c) renders this massing in a manner typical of the Northern Florida flora, suggesting the mass of a mossy Live Oak" (Steinfeld, 2019). In the images generated with StyleGAN, the "vertical columns define course features, such as camera direction and orientation, while horizontal rows define fine features, such as textures, colors, and lighting effects of each urban place" (Steinfeld, 2019 (Hesse, 2017).
The first application uses spatial data related to the surroundings of a generic house in order to determine the building outline. Information like location, the shape of the plot, and orientation have been used to infer the building footprint and general massing.
Chaillou trained a GAN model with Geographic Information System (GIS) data from the City of Boston (with plot shape and relative building footprint) so then the NN could predict generic results for any type of new plot of land.
Once the building outline is determined, NN is employed to predict the position of the windows (fenestration) and internal layout (  Visit the web version of this article to view interactive content.

What Model Is Best for Self-Organizing Floor Plans?
It is difficult, if not impossible, to say today with certainty which model is better for generating self-organizing floor plans. Related to it, the question also arises as to which one will be most used in the future. All of the methods seen here have their own respective pros and cons. For example, GANs are very successful at providing clean and sharp synthetic images (Karpathy et al., 2016), but their success is dependent on stability. The right balance needs to be struck between the generator and discriminator . Convolutional Neural Network (CNN) has been applied extensively and with relative success in the recognition of objects and parts of images We can see the initial input from the designer as a generic geo-referenced plot (7a), the final output generated by the GAN (7b), and its ground truth (7c): the original floor plan used to train the network (this process is applied to all plans in the set). Output floor plans and ground truth are often compared to evaluate the similarity between the computer generated and the original scheme. From: Stanislas Chaillou, https://developer.nvidia.com/blog/archigan-generativestack-apartment-building-design/?linkId=70968833. (LeCun, 2015, p. 439) with more recent application in facial recognition. More importantly, CNNs offer a great level of accuracy as the labeling can happen at the scale of pixels (LeCun, 2015, p. 439). However, they require large data sets and precise labeling to produce accurate results (LeCun, 2015, p. 440). NEAT is a powerful method to optimize a network by finding a minimized topology by searching for solutions with the minimum number of dimensions in each generation (Ibrahim et al., 2019, p. 111). NEAT has been used quite successfully in approaching real-time problems with user's input/action (Bird et al., 2019, p. 752). However, its performance significantly depends on the initially chosen topology (Ferner et al., 2017, p. 11) and the control of complexity within inner networks (Le Goff et al., 2020, p. 43).
So far, we have explored two main approaches to self-organizing floor plans. One based on the optimization of an initial configuration (this being an existing building floor plan or a hypothetical one based on a set of desired criteria), and the second focusing on the generation of a new configuration (based on prediction of new labels from a training set).
Although these two approaches can be combined (see, e.g., Chen at al., 2019) and share some logic (think of the generative aspects of both approaches), they are quite distinct in their objectives. In optimization problems using NEAT, the final spatial configuration is strongly conditioned by the initial configuration (i.e., existing floor plan), and by the ways in which inputs are given at each iteration of the Genetic Algorithm (GA). Optimization approaches in general yield more accurate results, for they provide a new spatial configuration that retains many of the original architectural features as a part of the topology (e.g., doors connected to walls as in the original plan). Conversely, with GANs the outcome is largely dependent on the training data set used, so factors like the size of data sets, initial data resolution, and quality of the data set are highly influential to achieve the desired results. New configurations generated through GANs are in general more innovative, as outcomes are often unexpected, yet they tend to be less accurate, as there is not an original floor plan used as starting point (and for comparison).
It is intuitive to see why in those cases where designers need an accurate and controlled new configuration that 'evolved' from the original one, optimization models through genetic algorithms are preferred. These allow designers to retain a good degree of accuracy in the position of key elements of the configuration of which movement in space requires careful consideration, including structural elements (load bearing walls, columns, beams, etc.), health and safety parts of the building (fire escape routes, fire compartments, etc.), and other building parts (shafts, ductwork, etc.).
By the same token, in those cases where designers are working with a larger degree of freedom (e.g., scoping out the possibilities of a new building or space, working with speculative design, etc.), generative models like GAN are effective in producing synthetized solutions. They allow designers to obtain often unforeseen configurations, generating an element of surprise and unpredictability in the design process.
The two approaches should be used in different instances with different aims, mostly depending on the stage of development of the project (e.g., conceptual versus development phases). As unsurprising as it may sound, the answer to the initial question: 'What model is best for self-organizing floor plans?' is that it depends on what designers are after in each particular phase of their project.
In general terms, however, we note that GANs (and their variants like StyleGAN or DCGAN) seem to be increasingly popular among computational designers and data scientists. This is most likely related to the fact that GANs can be trained with relatively modest data sets (e.g., Wallish's GAN Hadid project has around 8,000 images as a training set) and to the fact that image-based data sets are increasingly affordable and offer promising results. We should note that many of the GAN projects in this article have been developed using NVIDIA GPUs technologies (see Song et al., 2018) (e.g., Wallish, Chaillou, and Steinfeld's) that allow for relatively satisfactory results in small periods of time and with general-use machines. Most of these implementations of GANs are based on recent developments at NVIDIA for optimized image processing (Karras et al., 2017;Karras, 2020), as well as new tools that facilitate the training of data sets and the use of ML in general, including Pix2Pix (used by Chaillou and Wallish), and more generally TensorFlow, PyTorch, and so on(underpinning most of the projects using GANs).
GANs may not necessarily be the best solution for SOFPs, but their use is certainly increasing and they are attracting a growing number of designers who want to experiment with these models, given the growing number of publications and promising results.
If more designers are experimenting with methods like GANs and NNs in general as a part of their design explorations, access to the appropriate technology becomes a key part of this development. By access we refer to infrastructure (this includes hardware: CPUs, GPUs, or TPUs, computational power of general-use machines and software: open-access libraries and online resources) as well as skillset. For designers to be able to use relevant data science techniques and methods in their projects, not only do they need to understand the design problem and model it, but they also need to effectively implement their strategies. Thankfully, a growing number of libraries, Application Programming Interfaces (APIs), and platforms are now available to designers who can more easily access some of the data science methods mentioned in this article. The following section explores some of them, illustrating what designers can do and how.

How Are SOFPs Used (by Designers)?
The models presented in the previous sections illustrate cases where designers have a hybrid profile (data scientist/designers) and are able to work with the two aspects of these projects (data science/architecture) at the same time. Most of the techniques and tools used in the examples from Section 2 are not design tools per se. They are rather workflows generally used in computer graphics, computer vision and machine learning projects (for example, our NEAT project is coded entirely on Python and Pix2Pix is an image-to-image mapping tool). In parallel to this approach where designers use data science tools and methods, there is an important growing infrastructure of software, libraries, and interdisciplinary communities developed directly for designers. These tools are helping more and more to democratize important developments in data science into the field of design and architecture.
The libraries and software presented following represent a good section of the new tools for data science available today to designers and show the direction in which the development of computational tools is going, especially helping designers without a strong data science background to work with ML and NNs models.
The implementation of neural network methods (and machine learning in general) for computational designers has rapidly developed in the last decade. Firstly, there is the emergence of experimental approaches where individuals make promising attempts at running NNs through modeling and Computer-aided Design (CAD) platforms. These scripts are usually written in C# or Python, using developer kits within the CAD or parametric software (Khean et al., 2018). A good example are the ML tools within the Lunchbox Library for Grasshopper and Revit, and the beta version of the APIs in Grasshopper including Crow, a library with supervised learning through backpropagation networks and unsupervised learning through self-organizing maps. Figure 8 shows how the network is trained using an internalized database to associate basic bidimensional shapes with three-dimensional meshes representing objects (for example, fruits).
Visit the web version of this article to view interactive content.
Another good example is PlanBee, developed by Marco Juliani at the beginning of 2020. This library allows designers to use a Kohonen SOM algorithm to automatically organize blocks of activities in a given floor plan with pre-settings ( Figure 9). "In order to understand this, it helps to think of the floor plan as a field of voxels, each of which contains values for different computed metrics. Once the metrics are computed, each voxel corresponds to a multi-dimensional vector. The SOM basically reconciles the multi-dimensional vectors of the field of voxels with the features one specifies should belong to a programmatic mix (i.e. the list of activities in the building)" (Juliani, 2020). interface Dynamo and the option to integrate these tools with text-based programming languages, represents a significant shift in the design industry. While 5 years ago (and before), computational design tools (including ML libraries) were only add-ons to main programs and needed some tweaking by the designers in order to work (Daher et al., 2019), such tools are now fully integrated into the standard industry software, hence, they are largely available to any designer.
In addition, interoperable online tools to create workflows are increasingly available to designers as well. For example, RunwayML offers several ML tools that can be flexibly connected to other apps. There is also the platform COMPAS that offers an open framework for designers to share libraries and APIs that work outside of any specific software so to increase its interoperability. The fact that main design software companies are embracing generative design models, as well as NNs and ML methods, offering them as a part of their standard tools should be considered positively. The inclusion of these tools in the designer's toolbox should encourage designers to experiment and better understand the potentials and limitations of intelligent systems applied to architecture and spatial practices. By the same token, this also calls for more data scientists to be involved with the design industry, bringing their expertise in this field. More and more, there seems to be the need for 'bridging profiles' who can work across data science and design.
SOFPs represent a great and equally exciting challenge for designers (and those interested in automatic generation of spatial layouts) for two main reasons. The first is about design as a discipline and field of study. These new data-driven approaches require designers to reconsider their role within the design process, from designers of buildings through drawings, to designers of scripts that will generate buildings. The second reason relates to specialization and expertise. Today, designers need to work with methods that originate in data science, using tools and a logic that is naturally outside of their traditional (academic and professional) training. Designers need, therefore, to heavily rely on colleagues from the data science community, embracing a new level of multidisciplinarity.
However, a caveat should be considered. This has to do with the transparency and intelligibility of these tools, and the extent to which designers are fully aware of the mechanisms that underpin the tools. In other words, these new tools make it very easy to obtain results from complex operations without necessarily the need of being aware of the entire process that brings those results. For example, some tools make it very easy to compute a linear regression given a large data set, but designers still need to understand whether a linear regression was the best statistical method in that particular case. In using APIs and libraries, designers need to put a certain level of trust on the tool and their developers who made them, assuming (and accepting) that the implementations of those tools fits the needs of their project.
Generalizing, in the case of SOFPs it is becoming increasingly easy to use tools that generate spatial configurations or architecture in general (either with NNs, optimization algorithms, etc.). It is important that designers (and any other users of these tools) understand the logic and process behind these workflows to ensure a good level of control over the obtained results. In other words, to generate a self-organizing plan is becoming increasingly easy, but who guarantees the spatial quality of the results? After all, these spatial configurations are generally made to be built, that is, translated into physical spaces where people live, interact, and subjectively experience the world around them. One way of addressing this concern may be to look at the quality of the training data that we use for our models. The designer's oversight over the entire design, modeling, and implementation process is very important. However, the final results of our floor plans are directly dependent upon the initial data used in the model. This aspect has an important impact on the final design, often more significant than the algorithm used to generate it.

A Final (Design) Note
Most of the machine learning methods mentioned in this article (especially the ANN) require large data sets (in the order of thousands) to be trained. In the current attempts we seem to concentrate more on where to find good and reliable data sets (the larger the better), how to train our models, and how to achieve a reasonable level of accuracy in the training (with the resulting floor plans that have a good resemblance to traditionally designed plans). One aspect that is still underdeveloped (and perhaps underestimated) is the quality of the training data. We often use data sets that reflect existing buildings, with most of them characterized by average and often mediocre architectural qualities. Often, the rationale used in finding an initial data set is not about the quality of the plans (or their architecture, social qualities, etc.), but more about the quality of the data itself (are plans comparable to each other, drawn wellenough to be read without too many problems? are data easily available, how large is the data set? etc.). Our models are increasingly fast and accurate, yet they output the same quality that we give them as input. From a strictly design perspective, this is not good, as (good) designers aim at achieving a new level of architectural quality with every new project (usually learning from previous ones).
In the examples presented in this article, we identify some promising cases that used a data set that included high-quality architectural projects. For example, Erik Swahn's work included drawings in his training data set from websites that publish projects by international renown architects (plansofarchitecture and DeZeen). In his GAN Hadid, Sean Wallish trained his generative adversarial networks with photographs of buildings by Zaha Hadid Architects. Zaha Hadid is today regarded as one of the most talented recent architects. The two projects illustrate very clearly how GANs can be applied to architecture to produce new spatial configurations on the basis of highquality projects. In their work, however, we are not able to see any final, full-fledged floor plan (like perhaps we see in Chaillou's). In GAN Hadid in particular, we can see how very well-designed buildings are separated into their salient components (that can be geometries and architectural features that may stand out from the overall building silhouette) and reaggregated into a new composition. These new synthetized spatial configurations are made of the parts of many buildings, yet they do not look necessarily surprisingly new. One may argue that the designer here trained the GAN so well that the algorithm generated a design as the original architect would have done. If the main objective of an NN is to learn from a training data set to generate new configurations based on rules inferred, this would be good for architects in general terms. This would be relevant in Chaillou's case, where his method can be applied to high-quality floor plans to generate extraordinary solutions (note that Chaillou trained his GAN with generic floor plans from the City of Boston).
As we all know, a good design is a complex combination of several factors that may include consideration of the context, the users, normative constraints, experience from past projects, and so on. A good design also includes aspirational elements like the designer's ambitions for the project, their own way of seeing and interpreting the world through their ideology and culture, and the intention of improving people's lives (with an added quality of the physical environment, for example). These are aspects that, combined, encourage designers to improve their work every time. In this regard, one may argue that learning from existing projects (floor plans, spatial configurations, etc.) is simply not enough to achieve incremental architectural quality. By the same token, designers should try to integrate into their computational approaches to design elements ones that are so far not included and that are more subjective, aspirational, and striving for incremental improvement. At the same time, data scientists should develop ways in which ML methods can include the designer and users' subjective components that make design (in principle) better every time.
One of the next challenges is then how to produce SOFPs with perhaps smaller and more controlled data sets that generate an incremental level of spatial quality, including a certain degree of subjectivity (of the designer) and that ensure a good public acceptance of the resulting building. In other words, how to encode those important aspects that differentiate a generic (and sometimes mediocre) building from an outstanding architecture that can enrich people's lives?

Disclosure Statement
The authors have no disclosures to share for this manuscript.