adaptive fa[ca]de mayo 1, 2010Posted by christian saucedo in Technology research in surfaces.
Tags: T. research - Responsive surface, Technology research
Desarrollo tecnológico. superficie reactiva
Autor. Marilena Skavara
Investigación y desarrollo. Marilena Skavara
‘Adaptive Fa[ca]de’ explores the functional possibilities and performative characteristics of cellular automata (CA). In addition to the unique emergent behaviour of CA, a neural network enables a further computational layer to evolve CA behaviour to the context of its surrounding environment. Building upon the early work of Conway’s ‘Game of life’ and Stephen Wolfram’s extensive research on the wider implementation of CA, ‘Adaptive Fa[ca]de’ becomes a living adapting skin, constantly training itself from the history of its own errors and achievements.
Since Alan Turing and Boris Belousov, with no awareness of each other’s research, discovered that there is a set of simple rules that lay beneath almost every complex system in nature, the scientific, technological and ultimately architectural thinking has radically changed. We face “a New Kind of Science”, as Stephen Wolfram’s book on Cellular Automata is titled, where not all mathematically explained behaviours can be predicted – one that requires a more cognitive and analytical approach of the very elementary components that are responsible for complex, random-like, almost chaotic formations. The Newtonian design process characterized by distinct starting, middle and ending steps is replaced with a generative approach that goes beyond a linear procedure and instead utilizes loops, iterations and oscillating dynamics within the designed system. John Frazer and cybernetic pioneers Gordon Pask and Cedric Price introduced this radically different approach into the architectural discourse. In this approach the end design product is not known beforehand, only the set of attributes, rules and possibly goals to be achieved. Architectural configurations are treated as wholes, as interactive and performative dynamics, where time becomes inseparably intertwined with space, surrounding environment and people.
Adaptive fa[CA]de explores a wide spectrum of functional possibilities and performative characteristics of Cellular Automata (CA) . It is an endeavour to formulate a system based on simple CA rules that constantly alters its pattern by tilting its panels to seven possible angles to adapt to the ever-changing light levels of the environment, aiming to provide optimum light conditions to the interior of the building. With known orientation, location and surrounding buildings that cast different shadows on the façade and just one set of light sensors on the top of the building that measure where the sun is at any given time, a finite grid of cells/panels would then adapt to its environment by yielding different CA patterns. But how can a system, which obeys to simple rules on one hand but is intriguingly unpredictable on the other hand, be controlled to achieve explicit goals? The idea of identifying an analogy between the initial condition, the first row of the grid, and the overall average of the different states of the cells on the grid is not new. Stephen Wolfram and Christopher Langton have conducted extensive research explaining similar analogies found in some types of CA. The challenge was to further enrich this behaviour with a localized, regional control on the shaded areas of the building.
This is where artificial intelligence is utilized. By training an artificial Neural Network (NN) to understand the inherent emergent behaviour of specific Cellular Automata, the system could then yield patterns that respond successfully to the changes of the light levels over the course of the year. This would not only apply to an overall average of opened and closed panels of the whole grid, but also to regional averages of different shaded parts of the grid. Dealing with a deterministic bottom-up system with known initial condition and rules and unknown later-on behaviour asks whether and where such control could be obtained. Would it be in the initial condition itself, or in the regional optimums to be obtained?
To test, two different experiments were conducted. In the first, a large list of randomly generated initial conditions – that is, first rows of the grid – constitutes the possible answers the NN might give for several sun positions – values of the light sensors – over the course of the year. The system is then trained to learn how to match successful patterns with each set of sensor values. Using a complex CA rule with seven different states and a neighbourhood of three individuals, the randomly generated list is not exhaustive. This disadvantage, coupled with the inadequacy of the network to understand that every difference, however minute, in the initial configuration of such an emergent and deterministic system could lead to radically different patterns led to the second endeavour. A genetic algorithm was developed to ‘correct’ and train the NN so that it gradually learns how it should behave based on the error between the ideal local and overall averages and these of the produced pattern. Since no list was needed, the system is free to generate every possible pattern as long as it obeys to the given CA rule.
The outcome of implementing a genetic algorithm to train the NN was a robust system that, after a few hundred generations, learnt the inherent structural attributes of the CA rule. Well equipped with this knowledge, it could then adapt to its environment, providing optimum light levels to the interior and generating beautiful, kinetic patterns. The resulting system can be described as a living skin – unique for each possible environment – a regulator of the interior and the exterior and ultimately an ever-changing pattern with high aesthetic value.