Issuu

Post/anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments Machine Learning as Computational Design Strategy Beyond Standard optimization

2.1 Author: Alkhaja, Nour

Theory Seminar / RC3 Cluster / Living Architecture Theory tutor: Jordi Vivaldi Piera The Bartlett School of Architecture

Univeristy College London Bartlett School of Architecture 2017 - 2018

Post/anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments Machine Learning as Computational Design Strategy Beyond Standard optimization

Nour Alkhaja Student Number : 17035121 n.alkhaja.17@ucl.ac.uk Research Cluster 3 / Living Architecture Supervised by Jordi Vivaldi Piera

Submitted on the 13th of July of 2018

This essay is focused on debating generative systems in architecture. In particular, it studies the hybridized design processes of human designers and artificial intelligence through machine learning for the production of spatial assemblages in extreme environments. The research re-evaluates the current state of the production of architecture driven by computational form-finding as a strategy based on optimization, replacing it with an algorithmic alternative approach grounded on self-learning capacities. The results aims to provide inhabitable structures with large degrees of autonomy in order to deal with extreme scenarios. This alternative computational design strategy reconsiders the relation in between architects and users, placing them in the position of being participants of a flat symbiotic partnership with intelligent algorithms. Through a design proposal located on Mars and based on Machine Learning algorithm, the tiling assemblages are found to be not only responsive, adaptive, and reconfigurable, but also autonomous; surpassing thus the static and lineal state of conventional optimization.

Keywords: Algorithm, Computational, Logic, System, Aggregation, Tiling, Optimization, Responsive, Machine Learning, Artificial Intelligence.

abstract

Contents 1. INTRODUCTION.......................................................................................................................9 1.1 Design as a procedural thinking process...............................................................................10 1.2 Mechanizing the design thinking process with computers.................................................. 11-13 2. THE UNFOLD OF THE DIGITAL DESIGN THEORY...............................................................14 2.1 Design strategies in the computational age..........................................................................5-17 2.2 Optimization as a popular design strategy..........................................................................17-20 2.3 Genetic Algorithm & Multi Agent Systems............................................................................24 2.4 The rise of AI and Robotics.....................................................................................................24-26 3. WHY WE NEED TO GO BEYOND OPTIMIZATION............................................................26 4. AN ARCHITECTURE PRODUCED BY AI................................................................................ 27 5. INTELLIGENT ARCHITECTURE IN EXTREME ENVIRONMENTS.......................................28 5.1 Mars Exploration as a Possible Scenario for an Extreme Environment............................ 28-31 5.2 Computed spatial tiling.........................................................................................................31 5.3 Prioritized spatial tiling..........................................................................................................32 5.4 Predictive spatial tiling .........................................................................................................32-36 5.5 Reflections upon ArchiGo ....................................................................................................36 6. CONCLUSION........................................................................................................................37 7. BIBLIOGRAPHY........................................................................................................................37 8. LIST OF FIGURES....................................................................................................................38

table of contents

Post/anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments Machine Learning as Computational Design Strategy Beyond Standard optimization

1. Introduction In the book of “The second digital turn (Carpo 2017)”, Mario Carpo discussed the recent digital technologies that has allowed for complex forms to be brought to life by digital designers. Thanks to the advanced use of what he refers to as “the new science of big data” (Carpo 2017) algorithmic tools allow to “collect , store , and process increasing amount of data at decreasing costs” (Carpo 2017 ,p.18). According to Carpo, “Computational form-search “hasn’t only replaced known conventional mathematical means of form optimization, but also lead to a “post-human” (Carpo 2017, p.79) phase were big-data management is the norm. One of the examples discussed to prove this point is Achim Menges ICD/ITKE pavilion, whose shell designers were inspired by a biological model. The project uses optimization and fabrication tools in a ground-breaking fashion; the FEA structural optimization algorithm is implemented via continues optimization iterations ; when a satisfactory result is reached , the robotic arms are used to execute the design. Carpo wrote: “In this process of heuristic (not mathematical) optimization, every simulated model that was tried and discarded corresponded to a physical model that a traditional artisan would have made” (Carpo 2017, p.40). The previous description depicts features about the modern strategy used to produce architecture, optimization and computational form-finding; but we still haven’t escaped that dominant mentality where algorithms are still referred to ,in most architecture literature, as tools. As designers we have mechanized design processes in the same manner that modernism was influenced by the serialized identicality of the industrial age.

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

The

digital turn allowed for infinite variability, but only in the product and not in

the process. When considering the recent advancements in Artificial Intelligence, and in particular, machine learning, it’s worth to investigate if such technology could bring variability in the design process when coupled with the human thinking at varying degrees of dominance.

1.1 Design as a procedural thinking process It might not be an easy confrontation to delve into the intricate philosophical discussions about establishing a clear identification of the design process, but it could be helpful to depart from familiar territories;a designer offers his/her own speculation on how a functional space should be arranged and expressed in the physical world. As Dean Hawkes argued, such “speculation” (Hawks, cited in March 1976, p.465) usually entail an initial presumption about the design assignment , it then goes under repetitive procedures of analysis and modification until “a solution is found which approaches is close to the defined goals” ”(Hawks, cited in March 1976 , p.465) . Hawks

also spoke of “stereotypes” (Hawks, cited in March 1976) which are

established common models that include hints and suggestions about how different types of buildings ought to be addressed, the relationships between their parts and other criteria that help to guide the designer. Those stereotypes grow into more detail and specialization of sometimes standardized ranges of values with the evolution of the building profession. Those models have a subtle influence over the geometrical entities that are generated during the design process. The progress of this agenda necessitates the development of some sort of a language the designer uses to mediate his/her own viewpoint over how the design problem should be resolved ; this is usually referred to as design intention. This language is a one of a visual type that follow rules crafted by the designer and which the geometries of the design are subject to. The suggested procedural nature of the design process surfaced in numerous architecture literatures, and such views has further has been intensified with the evolution of computational tools. In his seminal work “Architecture Machine “(Negroponte 1970), Nicholas Negroponte understands design as a “step-by-step process of sharpening both comprehension and representation of one’s image of the problem” (Negroponte 1970, p.19).

Nour Alkhaja

1.2 Mechanizing the design thinking process with computers We can see so far how design is framed as an incremental procedure of geometrical variations within a relatively known space of possible transformational values, different combinations of those values will result in varying qualities. The transformational operations conducted on the initial shape/s are of a mathematical nature. Architecture therefore is deeply intertwined with mathematics, this inevitability has led to the definition of the design process as “the computation of shape information that is needed to guide fabrication or construction of an artifact” (Mitchell , cited in McCullough 1990 ) . The required skills to articulate such rules are further nurtured and developed by learning and acquiring new knowledge (Mitchell , cited in McCullough 1990 ), it also demands the designer to be highly aware of the context of the project and sensitive to any changes that may rise in it . A

strong correlation between design and computational procedures could be inferred

from the previous statements, since computation rely on algorithms which employ’s mathematical logic, inference, and abstract reasoning. Lionel March (1976) revealed the sequence of logical operations taking place inside the black box of “design”; debating design as a model of consecutive reoccurring reasoning operations (March 1976, p.18). The first is the “production inference “which seeks to establish “a novel composition” (March 1976, p.18) that “can only be inferred conditionally upon our state of knowledge and available evidence” (March 1976, p.19) , or in a simpler manner , a hypothesis , a coherent assumption about the characteristics of which the design should assume , and where we can introduce new values (March 1976 , p.18 ) . The second type of reasoning is the “Deductive” which” can be used to predict measures of expected performance by the application of further models and theories to the particular design proposal” (March 1976, p.19), it’s the part that handles prediction about performance but without determining any values, unlike the third type “Inductive reasoning “which actually provide an output value that can be used to rectify or alter the hypothesis. To further clarify we can refer to March’s “rational design procedures” (March 1976, p.19):

01.1

From a preliminary statement of required characteristics and

01.2

a presupposition, or protomodel,

We produce or describe, 01.3

the first design proposal

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

02.1

From design supposition and theory and

02.2

the first design proposal

We deduce, or predict, 02.3

the expected performance characteristics.

03.1

From the performance characteristics and

03.2

the first design proposal

We induce or evaluate, 03.3

other design possibilities or suppositions.

The cycle then begins again: 11.1

from a revised statement of characteristics and

11.2

further m or refined suppositions,

We produce, 11.3

the second design proposal, And so on.

The interrogation of values through algorithms is conducted via decision making, search trees, and probability algorithms to generate iterations of the design proposal. Yet, March clarified that “computer-simulated experiments would seem as an ideal vehicle for many – but by no means all – architectural science investigations” (March 1976, p.31). March’s depiction of the architectural investigation as science confirms the experimental aspect with multiple values, and computational tools seems a proper tool to iterate over unlimited number of such trials. But such tools suffer from a blind spot, which is the design context. When computational tools are used as tools, as problem solvers, a specifically customized search algorithm will tirelessly navigate through tons of data, looking for the proper value, but would never be able to piece the rest of the puzzle’s pieces together. It will fetch a value, but without a mean to understand why such value is appropriate or not, it would remain inferior in knowledge to the human designer. Nicholas Negroponte (1970) argued for a revolutionary concept of an intelligent machine capable of producing architecture, evolve and learn, equally to its human design counterpart. He prescribed such machines as they would need: “the adaptability of the human and the specify of the present-day machines. They must recognize general shifts in context as well as particular changes in need and desire” (Negroponte 1970, p.5). Other

tools such as CAD systems offered convenient platforms for mechanized

production of architecture such as drafting and modelling, but proved to fail as a design exploration tool. This was mainly due to how data is structured within their systems, geometries in CAD tools maintain constrained relationships between geometrical

Nour Alkhaja

elements when modified or altered (Mitchell , cited in McCullough 1990 , p.29 ). Therefor Mitchell called for a change: â€œIf

we really want CAD systems to support creative design exploration (not just

representation and analysis of completed designs) we will need to break away from simplistic, rigid notions of structure, and face up to the difficult problems of building systems that are sufficiently flexible and pluralistic in their handling of itâ€?( Mitchell , cited in McCullough 1990 , p.32).

Figure 1-1: The PDI (Production/Deduction/Induction)- model of the rational design process (March 1976, p.20).

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

2. The unflod of the digital deisgn theory “But, through Deleuze, it was a whole post-modern universe of thinking that offered itself, sometimes covertly or inadvertently, to the then nascent theory of digital design.” (Carpo 2015, p.9). Fast forward to the 90’s and we can observe how Mitchell prayers were answered. Designers at the dawn of the digital age strove to break away from the conformity of typical Euclidean geometry, much like how postmodern architects declared mutiny against the standardization of modernism. Perhaps one of the most notable achievements of post modernism, was its rejection to the cruel homogeneity which dominated most modernism works. Modernism -notably influenced by the industrialization thus repetitive copies for one ideal state – focused on purity and the production of a universal framework to design for what they called a “universal man” (Jencks 1978, p.24). Charles Jencks (1978) pointed out that designing for a non-existing three-meter monster was: “no doubt a logical necessity for architects and others who want to generalize a statistical average. They try to provide modern man with a mythic consciousness, with consistent patterns reminiscent of tribal societies, refined in their purity, full of tasteful ‘unity in variety’, and other such geometric harmonies “(Jencks 1978, p.25). Modern

architects therefor were unable to produce realistic works that respected

diversity, ignoring valuable factors of locality and variability. It was Gill Deleuze’s (1992) introduction to the concept of “Objectile” that brought to existence a new universe for digital designers to explore novel forms. The objectile is “an open-ended notation which allows for infinite parametric variations “(Carpo 2013, p.145), a native curvature that awaits the actuation from the designer to manifest how it reacts to different forces, producing endless variations. But it’s far away from any regular curvature. According to how Cache understands it, it’s either a convex or concave with an inflection, it “has the characteristics of a geometric undecidable, which works outwardly from its center.” “(Carpo 2013, p.148), paving the way for the parametric design. When reflecting upon the intellectual meta-human state at the dawn of the digital age, we can understand how cultural and technological forces pushed towards the notion of indeterminacy. A shift “from a universe of objects to one of relationships” (Frazer 2013, p.48). Designers and computer scientists developed an interest in simulating natural processes of evolution, through engendering a genotype (a script in computer science / an objectile in architecture), to test a proliferation of phenotypes (evolved scripts in computer science / phenotype geometries in architecture), producing unlimited breeds of unlimited variations.

Nour Alkhaja

Modernism was then keen to dedicate the idea of identical standardized copies, but it was overridden by “the capacity to design and mass-produce serial variations (or differentially) “(Carpo 2011, p.10), a character that “is specific to the present digital environment” (Carpo 2011, p.10). Untangling

the threads of the digital age story, will bring us to another turning

point “Authorship”. Architects always deemed themselves as the sole credited creator of their designs, but when they started to debate architectural forms as temporal instantiation, a scriptable probability, open for change and morphism, authorship became participatory. Once the original “scripted object” is released, the door opens for other designers to take over it and start experimenting, fine tuning its performance, customizing its characters: “the

idea of a generic, open-ended parametric, notation, … implies the possibility

that authorship may be split between more agents – on one side, the designers of the general function; on the other, its final customizers, or interactors “(Carpo 2011, p.8).

2.1 Design strategies in the age of computation

“Algorithmic logic is about the articulation of thoughts and a vague struggle to explore possibilities of existential emergence” (Terzidis 2006, p.40). With in a few decades a paradigm shift occurred. Authorship has become participatory not only among multiple designers but also with the algorithm itself. Design is no longer exclusively conceived as humanistic intuitive-based activity, but also a systemized rationalistic operation performed by a machine. Terzidis (2006) argued for the shared human-machine authorship over the results generated by algorithms: “Theoretically, ownership

of an idea is intrinsically connected to the predictability

of its outcome, that is, to its intellectual control. Therefore, in the absence of human control, the ownership of the algorithm processes must be credited to the device that produced it” (Terzidis 2006, p.21). To argue for how much a programmer, have control over its script would decide how each end of this symbiotic partnership -that started to crystalize - should be credited for. Computers were mainly used as tools due to their ability to reason about complex sets of information, that surpassed the ability of human designer to grasp and process. Different types of algorithms were put into use depending on the design aspect being addressed, some algorithms were designed to return deterministic values, such as testing for structural integrity, while other

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

Figure 2-1.1 : Objectile: Algorithmic Knots (Cache 2013, p.147).

Nour Alkhaja

algorithms were more of a probabilistic nature offering unexpected results which couldn’t have been known in advanced. For example, many systems were developed to assist in space/ activity allocation within a floor plan; where a designer introduces a set of required activities and constraints, the algorithm would return suggestions that might satisfy the defined criteria, but the suggestion fail when accommodated within the context. Such cases serve to further clarify the flaws of using typical algorithms as tools . The gap between what the user can achieve by using algorithms as tools and what they can offer when coupled with the ability to learn and evolve knowledge is mainly caused by the impotence of the human user who often fail to recognize the potential of AI technology. Architects have become aware of the fact that: “codified

information, such as standards, codes, specifications, or types, simply

serve the purpose of conforming to functional requirements, yet are not guarantors for a successful design solution” (Terzidis 2006, p.41). Optimization then is prone to failure without guidance from the designer to properly implementing it within the context. Most digital designers today lead an optimized computational form-search (Carpo 2017) design strategy to produce architecture, and are more inclined to giveaway partial authorship to the algorithm. It appears to be that the current stream in architecture encourages complementary relationship between the man and the machine, but more biased in favor of the human designer. Kostas Terzidis pointed that: “The

problem with this approach is that it does not allow thoughts to transcend

beyond the sphere of human understating. In fact, while it praises and celebrates the uniqueness and complexity of the human mind, it becomes also resistant to theories that point out the potential limitations of human mind” (Terzidis 2006, p.27). A mutual dialogue between both ends of this symbiosis will likely continue to morph in its boundaries, manifesting itself through competing digital design theories.

2.2 Optimization as popular design strategy The act of optimization aims to steer the design towards an ideal – or close to ideal- state of balance between conflicting set of criteria. By identifying parameters that represents desirable values or conditional relationships as constraints, a solution is sought either as an exact value or within a minimal acceptable range of error tolerance. Prior

to digital age, architects, sought to optimize topologies by referring to ideal

proportions, symmetry, and carrying out physical model simulations to test the quality of geometrical shapes performance under loads and stresses. Gothic architecture with

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

the tree-branching like structures are physical representation of the path of forces being transferred through structure. Gaudi used the hanging chain models to optimize arch shape to be used in archways and vaulting. Frei Otto soap film models were also an optimization of tensile structures behavior. High performance measures are then one of the most influential drivers for the design process. Optimization algorithms were adopted to digitally optimize design models. Depending on the stage at which optimization algorithms are implemented, we can distinguish tow strategies for the use of optimization in architectural design: either at an early stage of the design, where they are incorporated in solution generation, such as floor plans, topology or other quintessential aspects of the original concept. Or, after the concept has been developed to a relatively matured point, optimization is called to produce iterations of the concept to match a set of criteria. Both strategies rely on generating multiple model-instances, selecting the fittest iteration however is noted to the human designer whose also responsible for identifying the design objectives, setting the desired values to meet them, and expressing the former elements in a mathematical language in order to be understood by the algorithm. The type and complexity degree of input information should be also suitable to the design stage the designer wish to incorporate the optimization algorithm with. This requires the designer to possess exceptional analytical skills in terms of describing the problem. The degree of determinacy of the input data also indicate the type of the algorithm to be used. If the inputs are clearly defined, then a deterministic optimization algorithm would likely be chosen, such as FEA algorithms. While more random value solvers or stochastic optimization are used when the input data is vague or not known in advance and more of a suggestive nature than a specific value to be used. When human behavior is involved, it becomes challenging to articulate constraints in a numerical formula, in those cases we refer to either stochastic optimization or hybrid methods that would entail a degree of randomness when searching for solutions. One example of a stochastic optimization algorithm is the Evolutionary Structural Optimization or â€œESOâ€?. It was first introduced in 1992; the basic approach was about removing material from an element based on either the stress or the compress values, later a bi-directional algorithm was developed to include both addition and removal of materials for elements who act both under stress and compression. One of the first uses of that algorithm was in a four-story office building in Japan in Takatsuki, it was used to optimize the structure of the external faĂ§ade in a tow dimensional manner, resulting in a branching pattern with varying density where greater values were localized at regions of maximum loads and minimized at the opposite.

Nour Alkhaja

Figure 2-2.1 Genetic Algorithm: The classifier system responds to a set of environmental inputs and evaluates the relative success of that response. (Nagasaka, Frazer 1995 , p.79 ).

Figure 2-2.2: Emergent artificial life - evolving the rules (Graham, Frazer 1995, p.79 ,99).

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

In 2008, Arata Isozaki used the extended version of the ESO algorithm to process optimize a design of a branching tree-like structure for the National Convention Center in Qatar. In a reminder of gothic churches vaults, the shape was optimized: “for a particular set of site, material and structural loading criteria , this method archives the best , most efficient mechanically performing shape , using the least material possible”(Burry, Burry 2010 , p.130) . What can be inferred so far from the previous examples about the use of algorithms as optimization tools? Greg Lynn argued in a lecture about organic algorithms - a very interesting term as it transfers our perception of algorithms from passive calculus based tools into initializers or form engendering generators – that such practices are limiting us to optimizing to some perfect moment” (Lynn 2009), adding that “the whole idea of natural from shifted from looking for ideal shapes to looking for combination of information and generic from” (Lynn 2009). 2.3 Genetic algorithms and Multi-agent systems

“In

nature it is only the genetically coded information of form which evolves, but

selection is based on the expression of this coded information in the outward form of an organism. The codes are manufacturing instructions, but their precise expression is environmentally dependent” (Frazer 1995, p.14). Greg’s call to abandon the form-finding strategy in architecture, comes as an echo to the works of John Frazer who proposed an architectural model that is generative, adapting nature strategies of evolution and natural selection for proliferation and morphology. Those strategies also stood as a source of inspiration in computer science for Alan Turing who was intrigued by morphology in nature, the same goes for John Holland who replicated the processes of evolution and natural selection in the Genetic Algorithm. Bydrawing on a comparison between DNA strands in living cells, and the scripted functions within an algorithm, Johan Holland came up with the Genetic Algorithm (1970): “Software already has a genotype and phenotype, Holland recognized, there’s the code itself, and then there’s what the code actually does” (Johnson 2001). Architects

took notice of this technique that shifted their attention toward the

articulation of scripted processes – the genotype- rather than investigating the result of the code – the phenotype-.John

Frazer (1995) also clarified that an architectural

evolutionary model would also require a virtual hosting environment where the designer

Nour Alkhaja

would adjust parameters that can be used as fitness criteria, it would also imply a sort of intelligence, in the sense of evolving enhanced population over time. Essentially, GA

in computer science is recognized as a search procedure. It will

execute filtration within a population of scriptable entities based on the satisfaction of a fitness criteria by searching for successful candidates. Those members would have qualities allowed them to pass and would therefore be hybridized with their peers to further continue the enhancement of the population. An incremental bottom-up optimization process. After

many experiments with cybernetic models, Frazer was able to produce an

architectural evolutionary model called “Universal State Space Modeler” (Frazer 1995, p.83). A virtual self-similar space of “motes” -elementary logic entities- is constructed. The space is universal because it can map any space regardless of scale and complexity, each mote holds information about its state in terms of position, neighbors state, any assigned properties by the user, and a copy of coded rules that gets executed after evaluating the state of the mote with respect also to its neighbor’s state. The

mote data structure allows the user to define environment properties and

observe information being transferred from one mote to another, enhancing the direct correlation between environmental fluctuations and the evolved results influenced by the environment’s changes. An embedded GA code within the rules of each mote is responsible for evolving the rules thus producing adaptable assemblies with its environment. “Thus, the mapping process which leads from the code-script to the virtual model is essentially process-driven. It no longer presupposes a set of components … but a generative set of processes capable of producing the emergent forms” (Frazer 1995, p.90). This

strategy implies a learning aspect, as if the algorithm is evolving an innate

knowledge about which rules produce better results: ” It also means that the system as a whole can learn which rules are successful in developing and modifying form and so evolve new rules for improving its form-making ability” (Frazer 1995, p.88). In

Frazer’s universal model, rules are not imposed as final dictation for form

generation but rather a seed of a framework that allows autonomous interaction and evolution within the environment. Absolute top-down is absolute! Frazer’s experiments unfolded a new level of human-algorithm interaction in architecture. It demonstrated how the cultivation of algorithmic learning capacities lead to better unexpected results. It’s no longer an interaction between a user and a tool, but rather a continuous conversation between a user and a smart artificial interactor, who can respond, and evolve knowledge returning emergent results.

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

Manuel Delenda (2009) pointed out two key factors that stood behind the success of Genetic Algorithm: The first is variability: “heterogeneity and variation is what drives evolution” (Delenda 2009), without variation, an evolutionary system would lose the ability to adapt and a homogenous population with identical qualities would surface and deter because it would no longer be able to survive inventible change. The second key factor in Genetic Algorithm reasoning style is that there are multiple equilibrium states instead of one: “today’s

population thinkers realize that there is in fact multiple equilibria, that

multiple species are trapped in local optima, and there is no such thing as the ideal design” (Delenda 2009). A living evolving dynamic system is never the same at two consecutive moments, it’s always changing, evolving and assuming different equilibrium states with its unstable surrounding. Therefore, the mono-ideal state is absolute! However, GA was not the only algorithmic inspiration architects used to break away from the sterile strategy of the ideal finding. Multi-Agent systems provided a fertile experimental platform. Multi-agent systems rely on Object Oriented Programming (OOP) paradigm which allow for distributed decision making among multiple scripts: “Computational

agents are software entities that can sense their environment and

act on it. Artificial Intelligence of different levels of sophistication is used to endow agents with the capacity to reason about what they sense and decide how to act“(Delenda, Leach , Snooks 2010, p.49).They proved to be useful in modelling systems not only in architecture and urban design, but also in fields such as social science, mathematics, and economics. The ability to simulate the interaction of various sets of rules within the same environment has enabled us to observe the phenomena of emergence not only in a morphological sense but also in behavioral sense, since sometimes morphology is a result of multiple agents interacting. MicroImage is an experimental software that is about exploring the phenomena of emergence. A virtual environment inhabited by four basic types of agents or “organisms”, each with a special type of instructions on how to interact with the environment, would move and accelerate along a path, leaving visual trace: “The

structure generated through this process cannot be anticipated and evolve

through continual iterations involving alterations to the programs and exploring the changes through interacting with the software” (REAS, Leach, Snooks 2010, p.35). In MicroImage, the aggregation is generated via the movement of the agents, the rules of the agents dictate how they should react to the environment, the environment also responds in change to each organism/agent:” This method explores a balance

Nour Alkhaja

Figure 2-3.1: MicroImage Triptych (REAS, Leach, Snooks 2010, p.26).

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

between dynamic, generative software and controlled authorship” (REAS, Leach, Snooks 2010, p.35). 2.4 The rise of artificial intelligence and machine learning Implementing

artificial intelligence in architecture can then be done through various

ways such as Multi-Agent Systems or Genetic Algorithm. where an algorithm exhibits aspects of problem solving, logical deduction, learning, and knowledge acquisition. When tracing back the history of artificial intelligence as an independent field of study, it will become evident that the idea of a machine capable of human-like thinking process was debated in different scattered disciplines. Most AI related achievements focused on areas of problem solving, pattern recognition, classification, learning, communication in and between human and machines (Cybernetics), and logic inference. At some point, machines were inseparable from their programmed instructions, thus a large part of AI history was entangled with robotics. Perceptron (Rosenblatt 1957) was a machine equipped with artificial neural network to recognize. In his report about the Perceptron, Frank Rosenblatt wrote: “an increasing amount of attention has been focused on the feasibility of constructing a device processing such human-like functions of perception, recognition, concept formation, and the ability to generalize from experience” (Rosenblatt 1957, p.1). Alan Turing’s test “The imitation Game” reflects the ambition to reach a machine with a human like intelligence. Intelligence in humans can be defined as: “the ability of human beings to react to ever new challenges posed to them by the environment (natural or social) and to solve the problem involve” (Ratch, Ritcher, Stamatescu 1965, p.3). Such definition denotes a large proportion of intelligence to the skills of problem solving, except that such skills in humans endows other factors such as intuition, creativity and social empathy. Humans acquire knowledge through learning, but along the way, they associate information with other subjective projections that might be trivial to a machine, such as emotions, opinions, and intentions” (Ratch, Ritcher, Stamatescu 1965, p.9). But what if Robots were capable of social interaction? a question raised and debated extensively by social scientists. For example, Kathleen Richardson (2015) discussed the multiple opposing views, some advocating for a separation between humans and non-humans while others call for a post-human agency “rejecting the dualistic separation” (Richardson 2015, p.11). The debate involves the exchange of actions between different actants within a network of relationships, some social scientists called for the acceptance of a hybridized associations between humans and non-humans within a social model:

Nour Alkhaja

“Robots and AI systems show the limits of hybrid systems. Artificial intelligence and social robots try to create new hybrid forms where human-nonhuman attachments are recognized” (Richardson 2015, p.14). It could be valid

then to argue that the use of intelligent algorithms in architecture

could redefine the role of the architect, consequently altering the architecture being produced. The nature of agency between the two ends of this symbiosis would also influence the nature of this new architecture. Güvenç Özel (no date) argued that the current conventional use of robots – essentially algorithms – are a first

order cybernetic where there is a distinction between the

observer and the observed (the system), but when the observer becomes part of the system, the agency between the two becomes cybernetics of second order: “second-order system require context-aware robots, using advanced machine vision and an iterative operational code that can, like humans, readjust their actions based on unpredictable actors with nonrepetitive behaviours” (Özel, no date, p.100). Özel’s called for an intelligent living architecture conceived by a “context aware software and hardware that

would work together to produce artificial life and its spaces from

materials or bits “(Özel, no date, p.105). This living architecture assumes an equally able hardware which would be an actuator for the artificial intelligence. Perhaps then, AI would overcome the deficiency stigma some social scientists assigned to it due to its lack of physical embodiment. To understand why certain algorithm are intelligent, one must examine the process of information storage and retrieval. In conventional algorithms, each single data element is stored in a specific location within the memory, to retrieve it, the user must provide an exact symbol associated with the requested information and an exact retrieval address. While in Intelligent algorithms, information is stored via a learning process within a vector of values allocated at the memory. Artificial neural networks

ANN perceive

information as pattern or (input vector), the vector is associated with a set of synaptic weights that continues to be refined during multiple learning iterations to enhance accuracy when showed similar or identical patterns. The former description brings the AI a closer step to human intelligence, since humans can retrieve accurate results when presented with approximate or similar stimuli associated to a fact: “The continuous contact with the outside world delivers to the brain a constant stream of patterns to be stored in the brain and to be compared with already stored pieces of knowledge. Learning is equivalent to a reorganization of synaptic weights. In biological brains they are changed in a self-organizing process. In Artificial neural networks the same effect is achieved by special procedures” (Ratch, Ritcher, Stamatescu 1965, p.8).

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

If we pictured a dominance scale where human and AI at the tow opposite ends of it, we can start

assigning different types of human-algorithm agency in architecture to

varying degrees along that scale. At the far extreme end, we can find an intelligent selfproduced and self-motivated to survive, the architect then is a user.

3. Why we need to go beyond optimization ? In short, because we need better architecture. Ludger Hovestadt (2014), a professor at the renewed institute of ETH, referred to the current state of architecture as intoxicated with the idea of computers are simply machines not only to be used for optimization but also to optimized:” We’re stuck in the trap that is our idea of the computer as a machine” (Hovestadt, Kertzer 2014, p.12). No doubt that terminology reflects a position against certain topic, he argues that by conceiving computers as purely machines for problem solving to satisfy our obsession with optimization, we will keep running in the same loop: “Problems

no longer exist: there are no more problems with computers as machines.

Computers analyse the problems and the same computers then synthesize the corresponding solutions. One great big and very loud cybernetic feedback loop” (Hovestadt, Kertzer 2014, p.15). Eventually, we will run out of things to say. Almost 50 years ago, Nicholas Negorponte (1970) presented to the world his idea about an architecture machine that can engage in partnership with the human designer to produce architecture, beyond the limited scope of computer aided design: “One the one hand, in the context of computer aided design we are told to render into each other respective design functions and talents: man thinks and the machine calculates. On the other hand, in the context of automata studies we are told that: anything you can do, a machine can do better” (Negorponte 1970, p.25). Negroponte

then assumes an artificial design partner that is intellectually equal to its

human counterpart or even capable of surpassing it. Negroponte might have called his smart design partner a machine, but he denoted qualities to it that surpassed the repetitive mechanical context-blind deficiencies the word machine might suggest. He was not at a far distance from Hovestadt’s position, describing that such mechanical partner would be able to “discern changes in meaning brought about by changes in context” (Negorponte 1970, p.1). A context aware machine was then an imperative condition for it to be able to produce responsive structures, adaptability was implied as a target. Both the human and the artificial designer would be involved in a synthetic symbiosis of “tow intelligent species” (Negorponte 1970, p.1), collaborating

to produce ideas that could have never been

Nour Alkhaja

realized outside the premise of this partnership. Negroponte highlighted the motive behind such a machine, because “architects can’t handle large scale problems for they are too complex” (Negorponte 1970, p.3) and they often“ignore small scale problems for they are too particular and individual” Negorponte 1970, p.3),and they simply “act on the general” (Negorponte 1970, p.3). On the contrary to what opponents to Negroponte’s machine might assume about the inhuman nature such adventure could result of, he explained that such machines could in fact bring us closer to a more human architecture ; because of its ability to bridge the gap between the large-scale information -humans get caught up with and at many cases are unable to processand the fine granular level information they tend to ignore: “Because of this, an environmental humanism might be attainable in cooperation with the machines that have been thought to be inhuman devices but in fact are devices that can respond intelligently to the tiny, individual, constantly changing bits of information that reflect the identity of each urbanite as well as the coherence of the city” (Negorponte 1970, p.3). The artificial might succeed where the human has failed.

4. An architecture produced by AI Before questioning the nature of an architecture produced by artificial intelligence, we should question how it could be produced? The use of intelligent algorithms implies a system that handles the production of built spaces. It

also presumes that such system is capable of learning and producing

continuously modified structures. It would be valid then to expect highly responsive and adaptive structures when thinking of an architecture produced by AI. It would also be inevitable to assume that cybernetics would somehow be involved to simultaneously carry out the production in the physical space. Similar debate has been raised by Tristan Sterk (2006), who explored the nature of a system that would produce responsive architecture. First, he distinguished between two different models: Charles

Eastman’s Model that suggested “a machine led, systems approach to

control that is suitable for producing simple logical responses “(Sterk, Oosterhuis, Feireiss 2006, p.498), users would influence the system indirectly through interacting with the physical space. Yona Friedman’s model “provides an example of user led control” (Sterk, Oosterhuis, Feireiss 2006, p.498), where users through an interface can influence the system. Sterk

argued that the popularity of those models helped to promote the three

following ideas (Sterk, Oosterhuis, Feireiss 2006, p.494):

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

1) that architects design systems, not just buildings, 2) that feedback could be used as an architectural form generator, and 3) that the profession of architecture must respond to the changes that surrounded its practice. But Sterk proposed a hybrid model of the previous two. A system that exhibit adaptability because it can respond and learn from user’s interactions with the space while enabling them to directly control it. The context here would be understood by the system as a result of logical inference processes that accumulates input data from the environment, the user information, and the modifications carried out by the user to adjust the produced structure: “More

importantly the hybridized model can also be used to produce responses

that have adjustable response criteria, achieving this by using occupant interactions to build contextual models of the ways in which users occupy and manipulate space “(Sterk, Oosterhuis, Feireiss 2006, p.498). This model brings a different approach from the previously discussed approaches. GA and multi-agent systems suggests a bottomup approach, while Sterk’s system lies in the middle zone, it can be classified as a generative system that can observe, learn and generate architecture, while at the same time facilitate a top-down control.

5. Intelligent architecture in extreme environments 5.1 Mars exploration as a possible scenario for an extreme environment Impelled by curiosity and passion for exploration, humanity (and perhaps one day our AI-operated fiends) always manages to drive itself to new territories, off and on Earth. Upon confronting challenges brought by such brave endeavors, innovative ideas and solutions emerge to existence, ideas that sometimes force us to reevaluate the global social-cultural framework we imposed upon ourselves. Mars

hostile environment presents an opportunity to experiment with an intelligent

systems of architecture production. Due to the massive travel distance, and harsh unexpected climate, an intelligent system that can produce responsive adaptable structures seems appropriate.

Nour Alkhaja

Mars exploration scenarios hypothesized the existence of a habitat, where astronauts (users of the space) would need diverse set of spaces (sleeping, eating, growing food, lab experiments…). Such spaces would assume organization based on the context of the expedition (the target, the duration, the location, …), they should also have the possibility to be reconfigured to different arrangements to suite the changes in context. The reconfiguration should account for other factors derived from Mars environment (such as sun orientation, levels of radiation, ...) and factors that concerns the structure itself (stability, space connectivity, …). Reconfigurability then is highlighted as a major issue to be sought by the system.

Figure 4-1: Interactions Within The Hybridized Model Of Control (Sterk, Oosterhuis, Feireiss 2006, p.498).

Figure 5-2.1: Spatial Tiling System of ArchiGo (Alkhaja 2018).

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

Figure 5-2.2: A sample aggregation produced by the WFC algorithm (Alkhaja, Peljevic, Fereidouni, Tehrani 2018).

Figure 5-3.1: Weak tiles highlighted during ML training episode (Alkhaja, Peljevic, Fereidouni, Tehrani 2018).

Nour Alkhaja

Figure 5-4.1: An aggregation produced by AlphaGo agent (Alkhaja, Peljevic, Fereidouni, Tehrani 2018).

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

5.2 Spatially computed aggreagtions The proposed system uses a combination of algorithms to generate reconfigurable aggregations. The aggregations are an emergent arrangement of spatial tiles that offers various scenarios for space articulation when combined. The

system’s data structure is a self-similar 3D graph of space filling geometry –

truncated octahedrons -of which the spatial tiles can be mapped to. Each vertex within the graph has an address and stores information about its position, and the type of tile occupying it; also it can access information that the tile holds, such as the tile volume, the tile resource requirement, and a special assigned value called the tile weight. The

tiling system assumes a series of tiles designed to be contained within a

virtual vessel – the truncated octahedron – that the space would transfer through. The placement rules of each tile are what defines the emergent properties of the aggregation. In general, the placement is defined by a binary labeling system on the faces of the octahedron, since each tile allow connection to only certain faces of the octahedrons and restrict connection to others. A constraint solver algorithm called the Wave Function Collapse (WFC) is used to solve the propagation of the tiles with regard to the constraints defined for each tile. A function within the algorithm called the selector is responsible for the tile placement, for each vertex a domain of only few possible tiles is established based on the constraints of the neighboring available tiles, the selector would randomly choose a tile from that domain and place it. The framework has ingredients of tiles and a data structure that can be customized to produce aggregations with certain limited desirable qualities. Those aggregations that are produced by the WFC are tile assemblies solved for connectivity but can’t be described as adaptive because they lack the ability to respond to change in environment and to user defined rules. This where Machine learning steps in. 5.3 Prioritized spatial tiling In order to produce a reconfigurable adaptable responsive system, Machine Learning is used to build an intelligent selector that can select tiles with respect to certain defined rules.In Machine Learning, a scriptable entity called the “agent” is trained (iterate over a large number of learning episodes) to match certain state to a desired goal, this could be a set of values that needs to be calibrated to produce certain desired qualities but their value is not known in advanced. The algorithm uses the concept of reward and punishment to help guiding the learning towards the desired goals. The main difference between an optimization algorithm and ML algorithm is that the later seek to find a good approximation to a problem while the former seeks a higher level of accuracy and requires explicit input. ML algorithms seek to develop a model capable of solving multiple problems or inferring knowledge while the optimization algorithm exploit

Nour Alkhaja

search strategies to navigate the solutions space. The selector function is amended to prioritize the selection of certain tiles for the purpose of producing optimized aggregations, thus allowing for maximum adaptability, the produced aggregation would continually be updated to suit the context. This prioritization is done by favoring the tiles with the higher weight value within the domain selection. The tile weight is a virtual value assigned to the tile that indicate the likelihood of the tile to be placed when available in a domain, the value assignment is decided by the rules that can be placed either by the human designer, or by the algorithm that runs continues evaluations and autonomously adjust the tiles weights according to their fitness, or by both .At the first run, the selector would produce an aggregation with the WFC random selector, the aggregation would then be evaluated against a criterion such as the structural stability and the loose tiles would be assigned to a special list of weak tiles.The agent would lower the weight value for all the weak tiles, and its reward would be decided upon its ability to produce aggregation with fewer weak tiles. A pseudo code for the evaluation functions: for each ( tile in the tile set ) if ( tile displacement > displacement threshold ) { place the tile in the weak tiles list} else{ place the tile in the stable tiles list} If ( number of weak tiles > 2% of the total number of tiles) agent reward = -0.5 for each ( tile in the weak tiles list ) { tile weight -=1 } for each ( tile in the stable tiles list ) { tile weight +=1 } If ( number of weak tiles < 2% of the total number of tiles) { agent reward = 1 }

5.4 Predictive spatial tiling An aggregation produced by an ML agent is optimized to meet global criteria, such as required volume or sun exposure, stability, ...etc. But to have a granular level influence over the produced assembly, the tile placement should be considered with accordance to the neighboring tiles. The selector can favor one tile over the other not only by considering the global criteria but also local one, such as favoring certain tile assemblies due its spatial quality. The tile weight then is the accumulation of evaluating both local and global criteria. AlphaGo algorithm allows for predictive problem solving by back tracking through a search tree. The search tree is made up of matched states and outcomes, and continues to

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

Figure 5-5.1: Algorithms used in ArchiGo (Alkhaja, Peljevic, Fereidouni, Tehrani 2018).

Nour Alkhaja

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

AlphaGo algorithm allows for predictive problem solving by back tracking through a search tree. The search tree is made up of matched states and outcomes, and continues to grow and branch as the solver progress through learning iterations. The agent there for was hybridized with this algorithm to allow the selector to back track through a search tree before placing a tile, and choose the one that led to a better aggregation. 5.5 Reflections upon ArchiGo To evaluate ArchiGo, we can compare it to other similar experiments: Polyomino is a research agenda developed by Jose Sanchez (2014) at the University of Southern California. It’s focused on “missing topology mechanic” (Sanchez, Kretzer & Hovestadt 2014, p.126), where a system is composed of a kit of parts and “the missing element is not a number or parameter in the system but rather the topological diagram that puts them together” (Sanchez, Kretzer & Hovestadt 2014, p.126). Many similarities can be detected between ArchiGo and Polyomino. Both systems can be conceived as “missing topology mechanic” and they both use tiles to convey an aggregation. The basic data structure is similar except that ArchiGo implement an algorithmic approach to the tile selection via machine learning, while Polyomino relies on social agency to explore the aggregation. Polyomino sees the potential of a material becoming digital , thus the unites ta some point would become self-conscious and can regulate themselves (Sanchez, Kretzer & Hovestadt 2014, p.127).ArchiGo does not debate the nature of the used material and leaves the door open for future research to speculate on how system actuation can be executed . Sanchez coined the term “object–oriented design” to describe a “practice where the definition of identity comes before the process of formation” (Sanchez, Kretzer & Hovestadt 2014, p.128). A notion that can be detected in both systems. When comparing ArchiGo with John Frazer’s Universal Modeler (1992), we can find that tiles were also used in both models. Frazer’s model used tiles to create a skin that is supposed to enclose space, while ArchiGo uses spatial tiles that directly articulate a chain of connected spaces with various qualities, thus the articulation of space quality can be done through the tile design.Prefoliation rules in the universal modeler are evolved and enhanced via the Genetic Algorithm, thus the resulting aggregations are emergent and have emergent properties that can be influenced by altering the environment parameters.

Nour Alkhaja

ArchiGo

assumes a model that can be enhanced by implementing machine learning

with a greater degree of control over the algorithm. The algorithm in this case can act as autonomous self-correction system (bottom-up), and maintain the user control as an option(top-down) to influence the aggregation qualities.

Figure 5-5.1: . Polyomino ||: Dailymota (Li, Yu &CuiW 2014).

6. Conclusion The architect, the user and the algorithm, a symbiotic partnership: The project has investigated the production of architecture as a result of designing a system instead of a form, producing an intelligent universal strategy that handles the production of spatial aggregations. This strategy brings both the human designer and the artificial intelligence to a symbiotic partnership to produce adaptable responsive spaces that can continuously be adjusted responding to fluctuations in the context. Despite the fact that the final product is a virtual structure, it opens the door for further research to test its performance in the physical realm. What can be inferred from this experiment is that there are a number of ways to how AI can be thought of to produce architecture, and object-oriented design through machine learning is one of them. The approach choice counts on the type of the missing information the AI can be used to generate. The

symbiotic approach (between the user, the designer and the algorithm) re-

examines the boundaries of the architectâ€™s role, which would continue to change and morph as long as AI technology keeps moving forward.

Perhaps someday we could

reach a point where architecture is fully produced by an intelligent system, and the only way to influence it is by interacting with the system, the architect would become a user, or rather an informed user.

Post/Anthropocentric Algorithmic Systems for Tiling Assemblages in Extreme Environments

7. BIBLIOGRAPHY 1. Burry & Burry . 2010. The new mathematics of architecture , London : Thames & Hudson. 2. Carpo , M . 2011 . The alphabet and the algorithm , Massachusetts: The MIT Press . 3. Carpo , M. 2013. The digital turn in architecture 1992-2012 ,London : Wiley. 4. Carpo , M.2017 . The Second Digital Turn: Design Beyond Intelligence , Massachusetts: The MIT Press. 5. . C.E.B. REAS .2010. MicroImage/Process Compendium. In : Leach , N . Snooks , S . Swarm Intelligence: Architectures of Multi-Agent Systems ,2010 . 6. Colombia . 2009 . Manuel Delanda : Deleuze and the Use of the Genetic Algorithm in Architecture . Available form : https://youtu.be/50-d_J0hKz0 7. Delanda , M .2010 . Muti agent systems . In: Leach , N . Snooks , S . Swarm Intelligence: Architectures of Multi-Agent Systems ,2010 . 8. Frazer , J .1995 . Evolutionary architecture , London : Thames VII. 9. Hawks , D . 1976. Types, norms and habit in environmental design .In : March , L , . Architecture of from. Cambridge : Cambridge University Press , 1976 , pp 465-480. 10.Hovestadt ,L , Kertzer, M .2014. Alive :Advancements in Adaptive Architecture , Basel : Birkhauser. 11. Jencks , C.1978 . The language of post modern architecture , UK: Balding & Mansell Ltd. 12. Johnson , S. 2001. Emergence : The Connected Lives of Ants, Brains, Cities, and Software , Cambridge :The MIT Press. 13. March , L .1976. Architecture of from. Cambridge : Cambridge University Press. 14. Mitchell .W .1990. A new agenda for computer-aided architectural design. In : : Architectural Knowledge and Media in the Computer Era . (Cambridge :The MIT Press, 1990 , pp 1-16. 15. Negroponte , N . 1970 . The architecture machine . Cambridge : The MIT Press . 16. Ă–zel ,G .No date . Towards a Post Architecture . 17. Ratch ,U , Ritcher , M , Stamatescu , I .1965 . Intelligence and Artificial Intelligence: An Interdisciplinary Debate , Germany : Springer. 18. Rosenblatt , F . 1957 . The Perceptron :A Perceiving and Recognizing Automaton , Buffalo: Cornell Aeronautical Laboratory . 19. Richardson ,K.2015 . An Anthropology of Robot and AI , New York : Routledge 20. Sterk , T.2006 . Responsive Architecture: User-centered Interactions within the Hybridized Model of Control .In: Oosterhuis ,k , Feireiss ,L ,Game set and match II , Rotterdam : Episode Publishers .

Nour Alkhaja

21. Sanchez , J, 2014 . Polyomino: The missing topology mechanic .In: Hovestadt ,L , Kertzer, M , Alive :Advancements in Adaptive Architecture , Basel : Birkhauser. 22. Ted Talks . 2009 . Greg Lynn: How calculus is changing architecture . Available form : https://www.youtube.com/watch?v=DeyzUysMLy0 23. Terzidis , K .2006 . Algorithmic architecture , London : Architectural Press. 24. Willoughby , T .1976 . Searching for a good design solution . In : March , L, . Architecture of from. Cambridge : Cambridge University Press , 1976 , pp 379-407. interacting with the system, the architect would become a user, or rather an informed user. 8. LIST OF FIGUERS 1. Figure 1-1: The PDI (Production/Deduction/Induction)- model of the rational design process (March 1976, p.20). 2. Figure 2-1.1 : Objectile: Algorithmic Knots (Cache 2013, p.147). 3. Figure 2-2.1 Genetic Algorithm: The classifier system responds to a set of environmental inputs and evaluates the relative success of that response. (Nagasaka, Frazer 1995 , p.79 ). 4. Figure 2-2.2: Emergent artificial life - evolving the rules (Graham, Frazer 1995, p.79 ,99). 5.Figure 2-3.1: MicroImage Triptych (REAS, Leach, Snooks 2010, p.26). 6.Figure 4-1: Interactions Within The Hybridized Model Of Control (Sterk, Oosterhuis, Feireiss 2006, p.498). 7.Figure 5-2.1: Spatial Tiling System of ArchiGo (Alkhaja 2018). 8. Figure 5-2.2: A sample aggregation produced by the WFC algorithm (Alkhaja, Peljevic, Fereidouni, Tehrani 2018). 9. Figure 5-3.1: Weak tiles highlighted during ML training episode (Alkhaja, Peljevic, Fereidouni, Tehrani 2018). 10. Figure 5-4.1: An aggregation produced by AlphaGo agent (Alkhaja, Peljevic, Fereidouni, Tehrani 2018). 11.Figure 5-5.1: Algorithms used in ArchiGo (Alkhaja, Peljevic, Fereidouni, Tehrani 2018). 12.Figure 5-5.1: . Polyomino ||: Dailymota (Li, Yu &CuiW 2014).