Teaching an Artificial Intelligence for the Built Environment
As artificial intelligence permeates architecture and design, it is important to distinguish between different types of AI. In some cases, we know little about the problem at hand. In these situations, we rely on “unsupervised” machine learning to discover the underlying structure embedded within the problem. [1] In other cases, we may have information about the desired outcome. This can be used to guide and train the machine learning process. Implicit in this “supervised” machine learning is the idea that we must “teach” the AI. [1] This didactic role is an opportunity to forge a blended AI that is infused with human creative agency.
To train a machine, we need some notion of the desired outcome. In the context of architectural design, this can be established by specifying a bounded morphological design space. [2] Fortunately, architectural practice has a well established tradition of physical experimentation and form-finding. Antoni Gaudí, for instance, used hanging chains to study catenary arches. [3] Similarly, Frei Otto used soap film to approximate tensile structures. [5] Through the act of building and materialization, these architects developed an experience, a type of intuitive understanding of space. Why not utilize this knowledge to train an intelligence?
In a recent graduate research seminar, we explored these ideas in the context of studying minimal surfaces. While this topic has been well-studied in architecture, recent focus has been on using generalized physics engines that simulate soft-body dynamics using finite-element methods. [5]These simulators are easy to integrate with a digital design environment. However, over reliance on such tools may diminish the role of physical and material experimentation. Instead of tinkering with real objects, designers may simply run a physics simulation, in real-time, entirely within the digital context.
As an alternative, we test how physical experiments can be used to teach an AI. Specifically, we use an artificial neural network, which is a collection of artificial neurons, mutually connected with weighted synapses. Hidden neurons, the ones that connect inputs to outputs, seek to make sense of complicated, non-obvious and contextual information (“deep” learning refers to a neural network with multiple hidden layers). [6]
In our method, we start by building minimal surfaces according to specific design guidelines. See Fig. 1. These models are then digitized, producing training data. See Fig. 2. For each model, initial surface and anchor points are used to train the neural network (in the forward propagation process). Once the material is stretched into its final form, the deformed location of each point is used to optimize synapse weights (in the backward propagation process). See Fig. 3.
If we want to maintain a tectonic connection to the physical world, we must investigate how human knowledge can integrate with artificial intelligence. We propose that supervised learning is one potential method. This process leverages a tradition of physical experimentation as the means to teach an AI. By teaching machines about our built-environment, we may in fact learn more about our own design priorities.
End Notes
[1] Mitchell, Tom M. Machine learning. New York: McGraw-Hill Book Company, 1997.
[2] Krieg, O., Dierichs, K., Reichert, S., Schwinn, T. and Menges, A. ‘Performative architectural morphology: Robotically manufactured biomimetic finger-joined plate structures’. In Proceedings of the 29th eCAADE Conference, pp. 573–580. Ljubljana (Slovenia), 2011.
[3] Deutsch, Randy. Convergence: The Redesign of Design: AD Smart. West Sussex: John Wiley & Sons, 2017.
[4] Otto, Frei, and Ludwig Glaeser. The work of Frei Otto. New York: Museum of Modern Art; distributed by New York Graphic Society, Greenwich, Conn., 1972.
[5] Piker, Daniel. “Project Kangaroo — Live 3D Physics for Rhino/Grasshopper.” Space Symmetry Structure. March 10, 2018. https://spacesymmetrystructure.wordpress.com/2010/01/21/kangaroo/
[6] Miller, Steven. “Mind: How to Build a Neural Network.” March 12, 2018. https://stevenmiller888.github.io/mind-how-to-build-a-neural-network/
