Evolutionary Architecture
Biologic concept in Binary world
4 mins - reading time
Introduction
Evolution
Darwinian evolution is the idea of a blind process that adapts to the environment due to natural selection. The adaptation happens through the extinction of species that failed to learn how to adapt. Different types of species emerge through mutations, creating a huge variety. Mutations originate in genotypes, consisted of such a small entity as DNA, then they form phenotypes. This natural process has great creativity, level of adaptation, learning capabilities and robustness.
Architecture
Architecture has a need for constant change due to the rapid development of technologies, societal needs and climate change. That’s how the understanding of architecture as an active and responsive mechanism emerged or architectural machines concept. Flexible architecture that can grow, adapt, learn and evolve is vital for our constantly changing world full of uncertainty.
Evolutionary Architecture
Cedric Price, Fun Palace: Section showing potential use of interior spaces, 1963. Ink, coloured pencil and felt-tip pen on paper, 157 × 419 mm. Cedric Price fonds, Canadian Centre for Architecture.
Of course, for this purpose evolution, as a model that we borrow from nature, suits quite well. Even if we only draw inspiration from it, it can provide some working models to achieve a higher level of flexibility, adaptiveness and robustness.
In architecture, except for some attempts that use evolution as a generative force for an architectural form or its optimization, which is far from the point of this essay, there are not many examples. That is why, speculating about evolutionary architecture, I will consider different models and examples from other disciplines that however can be implemented in architectural machines.
Binary World
Our civilization is very technologically oriented. The computational power is growing with every year, we are surrounded by gadgets, electronic devices and other machines all the time. Technologies are deeply intertwined with our culture. Philippe Morel argues that everything that matters now are patterns and their recognition. Patterns consisted of 0 and 1, the logic brought to life through the wide use of capacitors.[1] However, even with this simple logic binary machines have managed to reach a certain level of intelligence.
Ray Kurzweil in The Age of Spiritual Machines asks the question if an intelligence created by another intelligence can surpass it. He examines evolution as the process that created us. He considers it remarkably complex and amazingly simple at the same time. Building its natural machines atom by atom with amino acid chains, molecules consist of atoms and can perform the transcription, error detection and correction operations. However, evolution is a very sloppy programmer because most of the code does not compute and appear to be useless (most of the sequences do not end up in proteins).
Another weakness from his perspective is that evolution is very slow. The speed for an intelligent process matters and evolution took billions of years to develop its marvellous designs. He argues that human intelligence, as a product of evolution, is more intelligent than its creator and the machines we produce will surpass us. [2]
The beauty of computation is that it can indeed make Evolutionary Architecture faster than natural evolution and far more controllable. However, can it due to its simple binary nature reach the same level of complexity or at least a comparable one? Machines, consisted of predesigned elements and fixed parts, are they able to evolve? Is it possible for them to adapt to something that is not foreseen by their designer? Having some failures or loss of some parts, can they manage to realise it and continue the performance adjusting to this new situation?
Evolutionary Models
After Jon von Neumann’s concept of Cellular Automata, it became clear that complexity can emerge from a simple ruleset. This, in turn, gave some confidence in trying to study and construct different evolutionary models. They all came from the Alife field, which started from modelling biology to simulate and prove some hypotheses. But what is interesting is that these models succeeded to achieve emergent unpredictable behaviours operating inside of the binary brain.
The model I would like to start with is Tierra, a computer simulation developed by Thomas S. Ray at the beginning of the 1990s. The model represented a computer with programs that compete for resources of the processing unit. Each program tried to copy itself elsewhere in the memory to use more resources.
Figure 1: Tierra by Thomas S. Ray. Early 1990s. Red – hosts, Yellow – parasites. Blue – immune hosts. “TIERRA.” Tierra.
Figure2: Tierra by Thomas S. Ray. Early 1990s. Red – hosts, Yellow – parasites. Blue – immune hosts. “TIERRA.” Tierra.
Once the memory was full of copies of the original seed program, a very interesting behaviour appeared. Some programs lost in this sort of natural selection failing to run and were removed. Others started to optimize themselves and became smaller because of a better position in scheduling policy for a smaller size. Some programs went further and shrank to the detriment of its self-replicating capability. That is how one type of mutation originated – the parasites. They could not copy themselves, but they could trick a larger program to do it. Then hyper-parasites emerged and other social programs that could only reproduce with the help of others. (Figure 1, 2)
The model of Tom Ray caused a great excitement because it showed that a simple ruleset can introduce very complex behaviours and unpredictable results, even the whole ecology. Instead of designing intelligence, intelligence can emerge from a set up playground.[3]
A few years later Karl Sims presented his Evolving Virtual Creatures, a system where creatures evolved through their adaptation in different environments. In contrast with Ray’s model, his system contained genotype to phenotype translation. (Figure 3) Each creature consisted of different size boxes. Its genotype had graphs with properties for body parts. The phenotype was derived from a genotype and looked like boxes with a neural network, some of them also had sensors and actuators.
Then, each creature was placed in a specific environment or task in a simulated three-dimensional world with Newtonian physics. Karl Sims had a hundred creatures simulated in a generation and they all were evaluated how well they fit for a particular problem. The variety came from mutations each generation had. Eventually, they tended to evolve towards results that indeed can be observed in nature (a snake for a swimming group for example). [4] (Figure 4)
In this model, even having a quite defined fitness function, it shows a great result in terms of adaptation to the environment through learning without any human intervention. Emerging from very simple input it builds complexity and shows extremely interesting results.
Figure 3: Evolving Virtual Creatures by Karl Sims. 1994. Designed examples of genotype graphs and corresponding creature morphologies. Proceedings of the 21st Annual Conference on Computer Graphics and Interactive Techniques - SIGGRAPH
Figure 4: Evolving Virtual Creatures by Karl Sims. 1994. Creatures evolved for swimming. Proceedings of the 21st Annual Conference on Computer Graphics and Interactive Techniques - SIGGRAPH
Resilience
Karl Sims with his project can be seen as a predecessor of a new field – Evolutionary Robotics. Basically, instead of working purely in simulation as previously reviewed models, evolutionary robotics takes a step further coupling simulation with a real robot.
What makes this field unique is the concept of embodied intelligence it follows and a wide use of evolutionary algorithms. The concept of embodied intelligence is an alternative view in a robotic field where the robot, the environment and their interaction are addressed as a whole. Evolutionary computing is used for optimization, modelling and design problems, and it was proved to be a very successful method for these challenges.
So Evolutionary Robotics applies the principles of natural evolution (evolutionary algorithms) for a holistic approach for engineering, where problems of designing the morphology of the robot, its sensory apparatus and controlling mechanism are addressed simultaneously and with a direct interdependency between the robot and the environment.[5]
The most well-known and influential representative of the Evolutionary Robotics field is Josh Bongard. He argues that robots need a robust performance under uncertainty. Bongard draws attention to the problem that in case of unexpected damage most machines fail, which showcase a low level of robustness and adaptiveness.
Figure 5: The logic of resilient Machines project’s robot by Josh Bongard, Victor Zykov, Hod Lipson (2006) “Resilient Machines Through Continuous Self-Modeling.”
In his project Resilient Machines, developed with Victor Zykov and Hod Lipson, the robot proved to recover from damage or loss of its parts through continuous self-modelling. This robot, called Starfish, is a four-legged robot that has two joints with motors in each leg, joint angle sensors in each joint and two tilt sensors in the main body, sensing how much it tilts in left-right and front-back directions.
So the robot does not have direct information about its look and the environment. However, Starfish is able to simulate in an embedded physics engine and has information about the parts of its body. This is enough to construct itself comparing the simulation with the data from the sensors.
Starfish uses a physics engine to evolve controllers in simulation with the help of evolutionary algorithms first and then examines some of those in the real world, continuously updating the physics engine. When the robot has damage or loss of its parts, the simulation stops matching the reality. For this reason, the robot is forced to adapt the simulation to reflect these changes, evolving a new controller. After the loss of its leg (Figure 5), Starfish was able to construct new models that were accurate enough to produce compensating behaviours, enabling it to move forward.[6]
This continuous learning strategy with minimum embedded information achieves a high level of adaptation to the damage recovery, new environment, completely unpredictable situations which is a valuable quality for Evolutionary Architecture.
New Senses
Another extremely valuable evolution for architectural machines is the ability to acquire emergent sensory capabilities that are needed in a changing environment. The advantage of biological elements is that they can self-regulate while binary machines usually have fixed parts and direct specification.[7]
The machine has to grow its own sensors that are not defined in an initial design for true flexibility and adaptiveness. This problem was addressed by Gordon Pask in his series of electrochemical experiments in the 1950s.
Figure 6:Electrochemical experiments by Gordon Pask and Stafford Beer (late 1950s). Schematic indicating the relationship between the electrode array and the ferrous sulphate medium. Cariani, Peter. “To Evolve an Ear. Epistemological Implications of Gordon Pask's Electrochemical Devices.”
The system he developed with Stafford Beer consisted of metallic iron threads, forming between small platinum electrodes, inserted in a ferrous sulphate solution and connected to an electrical source with a limited current. A high current passing through electrodes was resulting in the formation of dendritic metallic threads. If no current passed through a thread, it was returning to the acid solution. This force is a primary source of formation, but neighbouring branches also take part by physically creating an obstacle.[8] Because the source of current is limited, threads start to compete for resources, which is a similar principle with Ray’s system.
Figure 7: Electrochemical experiments by Gordon Pask and Stafford Beer (late 1950s). Cariani, Peter. “To Evolve an Ear. Epistemological Implications of Gordon Pask's Electrochemical Devices.”
The main novelty of this assemblage was that the threads could be steered to develop sensitivity for some types of perturbations through a rewarding system.
"We have made an ear and we have made a magnetic receptor. The ear can discriminate two frequencies, one of the order of fifty cycles per second and the other on the order of one hundred cycles per second. The ear, incidentally, looks rather like an ear. It is a gap in the thread structure in which you have fibrils which resonate at the excitation frequency." [9]
A similar strategy can be implemented in architectural machines to automatically adapt, growing new senses to vibrations, sound, temperature, chemical elements or magnetic fields for example.
Evolving Through Human-Machine Interaction
Gordon Pask also realised the importance of human-machine interaction and its use for emergent complexity through looping his systems. Architecture or architectural machines can evolve through learning from humans for a better social performance and more flexibility to different kinds of changes.
Minimaforms (Theodore and Stephen Spyropoulos), exhibited the evolution of a robot caused by a reaction to human engagement in an AI robotic installation ‘Petting Zoo’. The robot has the form of suspended robotic arms that change its behaviour reacting to human contact and each other. Through computer vision, the machine is aware of people around, when they approach or touch the installation. This data is processed and patterns between the robot’s actions and human responses are identified, which lets the behaviour of the robot evolve. The Pets interact with humans through colour and brightness, movement and sound. By colour, they can indicate their emotional state which can be anger or fear, boredom or a playful mode. [10] Making the human-machine interaction emotive and sensorial also stimulates the participation of visitors.
Figure 8: Barbican Petting Zoo by Minimaforms, 2013 Petting Zoo, Barbican Centre, London.
By the end of two exhibitions (in FRAC Centre and in Barbican in 2013) each robotic pet developed personalities according to the environment, context and their experience. [11] They evolved through time, learning from this social engagement.
Evolutionary Architecture or architecture in general must have the last aspect which is evolving through response to humans. It must exhibit the active behaviour and the intelligence we are able to provide through engineering binary architectural machines. Minimaforms shares this vision and treats the installation as an explorative tool for use in architecture.
“"What if a home took on the characteristics of its inhabitants?" he [Theodore Spyropoulos] says. "What if a house could influence its own self-organisation within larger clusters of apartments?"”[12]
Conclusion
Evolution as a phenomenon cannot be directly translated to our binary machines. But it can be used as a source of inspiration for adjusting some of its principles to machinic language. Evolutionary architecture, viewed here as architectural machines, can exhibit a high level of adaptation, flexibility and robustness despite its simple binary logic. Examples, viewed in this essay, showcase the potential ability to design machines with a sufficient level of complexity and adaptiveness to unforeseen changes in the environment and in society.
Evolutionary models from the Alife field by Thomas Ray and Karl Sims prove the ability of machines to have a complex ecology through simple rule-set and adaptation to the environment in the simulation field. Evolutionary robotics went further and proved these theories on real robots. Josh Bongard showcased the mechanic robot that can recognise its damage and reprogram itself to continue its performance, which is a great strategy for resilient and robust architectural machines. Of course, during evolution species were changing a lot, acquiring new organs of perception, new ways of breathing and other components of the body. The case, provided by Gordon Pask and his electrochemical experiments, proves that certain setup, that is implementable in binary architectural machines as well, can acquire new senses that were completely missed in the initial design. Furthermore, Evolutionary Architecture can also evolve through social engagement and even develop its personality which was showcased through the work of Minimaforms.
Evolutionary Architectural Machines must merge the strategies exhibited here (or similar ones) to achieve strong interdependency with the environment and the people inhabiting it. It is a machinic version of a natural organism that can grow senses, adapt to unexpected situations, interact and evolve through time without having drawbacks of natural evolution such as uncontrollable behaviour and slow speed of change. It is very important for architecture to exhibit these features in order to be relevant in today’s dynamically changing world with a huge number of unpredicted situations and catastrophes.
[1] Philippe Morel, Architecture: Patterns, Databases and the Giant Global Graph. (AA PhD Seminar: London, 19.02.2020).
[2] Raymond Kurzweil, The Age of Spiritual Machines: When Computers Exceed Human Intelligence (New York: Penguin Books, 2000), 16.
[3] Rodney A. Brooks, Flesh and Machines How Robots Will Change Us (New York, NY: Vintage Books, 2003), 181-183.
[4] Rodney A. Brooks, Flesh and Machines How Robots Will Change Us (New York, NY: Vintage Books, 2003), 181-183.
[5] Stephane Doncieux et al., “Evolutionary Robotics: What, Why, and Where To,” Frontiers in Robotics and AI 2 (March 2015), https://doi.org/10.3389/frobt.2015.00004, 1-2.
[6] J. Bongard, V. Zykov, and H. Lipson, “Resilient Machines Through Continuous Self-Modeling,” Science 314, no. 5802 (2006), https://doi.org/10.1126/science.1133687, 1118-1121.
[7] Peter Cariani, “To Evolve an Ear. Epistemological Implications of Gordon Pask's Electrochemical Devices,” Systems Research 10, no. 3 (2007): pp. 19-33, https://doi.org/10.1002/sres.3850100305, 1.
[8] Jon Bird and Ezequiel Di Paolo, “Gordon Pask His Maverick Machines,” The Mechanical Mind in History, August 2008, pp. 185-211, https://doi.org/10.7551/mitpress/9780262083775.003.0008, 201-202.
[9] Gordon Pask, “The Natural History of Networks” in Self-Organzing Systems, ed. In M. C. Yovits and S. Cameron (New York: Pergamon Press, 1960), 261.
[10] Kaamil Ahmed, “A Digital Petting Zoo Is a Template for Homes That Learn to Adapt to Their Surroundings,” WIRED (WIRED UK, October 4, 2017), https://www.wired.co.uk/article/dont-feed-the-robots.
[11] “Petting Zoo, Barbican Centre, London,” Minimaforms, accessed March 29, 2020, http://minimaforms.com/.
[12] Kaamil Ahmed, “A Digital Petting Zoo Is a Template for Homes That Learn to Adapt to Their Surroundings,” WIRED (WIRED UK, October 4, 2017), https://www.wired.co.uk/article/dont-feed-the-robots.
Bibliography
Ahmed, Kaamil. “A Digital Petting Zoo Is a Template for Homes That Learn to Adapt to Their Surroundings.” WIRED. WIRED UK, October 4, 2017. https://www.wired.co.uk/article/dont-feed-the-robots.
Bird, Jon, and Ezequiel Di Paolo. “Gordon Pask His Maverick Machines.” The Mechanical Mind in History, 2008, 185–211. https://doi.org/10.7551/mitpress/9780262083775.003.0008.
Bongard, J., V. Zykov, and H. Lipson. “Resilient Machines Through Continuous Self-Modelling.” Science 314, no. 5802 (2006): 1118–21. https://doi.org/10.1126/science.1133687.
Brooks, Rodney A. Flesh and Machines How Robots Will Change Us. New York, NY: Vintage Books, 2003.
Cariani, Peter. “To Evolve an Ear. Epistemological Implications of Gordon Pask's Electrochemical Devices.” Systems Research 10, no. 3 (2007): 19–33. https://doi.org/10.1002/sres.3850100305.
Doncieux, Stephane, Nicolas Bredeche, Jean-Baptiste Mouret, and Agoston E. (Gusz) Eiben. “Evolutionary Robotics: What, Why, and Where To.” Frontiers in Robotics and AI 2 (2015). https://doi.org/10.3389/frobt.2015.00004.
Kurzweil, Raymond. The Age of Spiritual Machines: When Computers Exceed Human Intelligence. New York: Penguin Books, 2000.
Morel, Philippe. Architecture: Patterns, Databases and the Giant Global Graph. AA PhD Seminar: London, 19.02.2020.
Nygaard, Tonnes F., Charles P. Martin, Jim Torresen, and Kyrre Glette. “Self-Modifying Morphology Experiments with DyRET: Dynamic Robot for Embodied Testing.” 2019 International Conference on Robotics and Automation (ICRA), 2019. https://doi.org/10.1109/icra.2019.8793663.
Nygaard, Tonnes F., Charles P. Martin, Eivind Samuelsen, Jim Torresen, and Kyrre Glette. “Real-World Evolution Adapts Robot Morphology and Control to Hardware Limitations.” Proceedings of the Genetic and Evolutionary Computation Conference on - GECCO '18, 2018. https://doi.org/10.1145/3205455.3205567.
Nygaard, Tonnes F., Charles P. Martin, Jim Torresen, and Kyrre Glette. “Evolving Robots on Easy Mode: Towards a Variable Complexity Controller for Quadrupeds.” Applications of Evolutionary Computation Lecture Notes in Computer Science, 2019, 616–32. https://doi.org/10.1007/978-3-030-16692-2_41.
“Petting Zoo FRAC Centre.” Minimaforms. Accessed March 29, 2020. http://minimaforms.com/.
“Petting Zoo Prototype.” Minimaforms. Accessed March 29, 2020. http://minimaforms.com/.
“Petting Zoo, Barbican Centre, London.” Minimaforms. Accessed March 29, 2020. http://minimaforms.com/.
Sims, Karl. “Evolving Virtual Creatures.” Proceedings of the 21st Annual Conference on Computer Graphics and Interactive Techniques - SIGGRAPH '94, 1994. https://doi.org/10.1145/192161.192167.
Yovits, M. C., and Scott Cameron. Self-Organising Systems. New York: Pergamon Press, 1960.
Follow for more