Academia.eduAcademia.edu

Figure 45 – uploaded by Nizam Onur Sönmez

See full PDFdownloadDownload figure
FIGURE 4.21 Comparison chart for average performances of different version combinations.

Figure 4 21 Comparison chart for average performances of different version combinations.

Related Figures (150)

FIGURE 1.1 Three main methodological threads of the study. FIGURE 1.2 Basic research instruments of the study.
FIGURE 1.1 Three main methodological threads of the study. FIGURE 1.2 Basic research instruments of the study.
FIGURE 2.1 Problem-solving task used in the experimentation by Alexiou, Zamenopoulos and Johnson [2009]. FIGURE 2.2 Design task used in the experimentation by Alexiou, Zamenopoulos, and Johnson [2009].
FIGURE 2.1 Problem-solving task used in the experimentation by Alexiou, Zamenopoulos and Johnson [2009]. FIGURE 2.2 Design task used in the experimentation by Alexiou, Zamenopoulos, and Johnson [2009].
in the problem-solving task, a criterion for deciding the termination of the task is given, as well as a definition of legal moves. The problem itself is well-defined, the legal moves are known and here is a unique set of equivalent solutions (Figure 2.1). In contrast, in the design task, there s no predetermined final state or criterion for deciding the termination of the task. The task is )pen-ended, and requires defining the problem as well as the solution space. It requires the creation and interpretation of a set of moves, as well as the creation of criteria for evaluating the solutions (Figure 2.2).
in the problem-solving task, a criterion for deciding the termination of the task is given, as well as a definition of legal moves. The problem itself is well-defined, the legal moves are known and here is a unique set of equivalent solutions (Figure 2.1). In contrast, in the design task, there s no predetermined final state or criterion for deciding the termination of the task. The task is )pen-ended, and requires defining the problem as well as the solution space. It requires the creation and interpretation of a set of moves, as well as the creation of criteria for evaluating the solutions (Figure 2.2).
FIGURE 2.3 The co-evolution model of Maher, Poon, and Boulanger [1996].
FIGURE 2.3 The co-evolution model of Maher, Poon, and Boulanger [1996].
FIGURE 2.4 Analysis - synthesis - evaluation model of Lawson [2005, p. 49].
FIGURE 2.4 Analysis - synthesis - evaluation model of Lawson [2005, p. 49].
FIGURE 2.5 The model of design problems via constraints [Lawson, 2005, p. 106].
FIGURE 2.5 The model of design problems via constraints [Lawson, 2005, p. 106].
FIGURE 2.6 Model of design activities [Lawson and Dorst, 2009, p. 51). The second model (Figure 2.6) describes design process through groups of activities and skills that are commonly observed in successful design. Main activities include formulating, representing, moving, evaluating, and managing (Lawson, 2005; Lawson and Dorst, 2009, p. 51). As such, this model can be seen as a guide for determining types of design moves and corresponding types of operators.
FIGURE 2.6 Model of design activities [Lawson and Dorst, 2009, p. 51). The second model (Figure 2.6) describes design process through groups of activities and skills that are commonly observed in successful design. Main activities include formulating, representing, moving, evaluating, and managing (Lawson, 2005; Lawson and Dorst, 2009, p. 51). As such, this model can be seen as a guide for determining types of design moves and corresponding types of operators.
FIGURE 2.7 Levels of design activities [Lawson and Dorst, 2009, p. 61]. The third model concerns the levels on which the above activities take place (Figure 2.7). These levels are ‘project’, ‘process’, ‘practice’, and ‘profession’. Project level concerns a particular design problem, while process is a more general collection of methods or ways of working of a designer. Through their reflective moments, designers learn from their projects and develop their own approaches to design problems. This is crucial to developing a strategic view of design and possibly a distinctive style of designing. This brings us into the practice level, which comprises the style, assumed roles, participation, and partnership issues for a designer's professional practice.
FIGURE 2.7 Levels of design activities [Lawson and Dorst, 2009, p. 61]. The third model concerns the levels on which the above activities take place (Figure 2.7). These levels are ‘project’, ‘process’, ‘practice’, and ‘profession’. Project level concerns a particular design problem, while process is a more general collection of methods or ways of working of a designer. Through their reflective moments, designers learn from their projects and develop their own approaches to design problems. This is crucial to developing a strategic view of design and possibly a distinctive style of designing. This brings us into the practice level, which comprises the style, assumed roles, participation, and partnership issues for a designer's professional practice.
FIGURE 2.9 General scheme of generate and test algorithms.
FIGURE 2.9 General scheme of generate and test algorithms.
FIGURE 2.10 Local search by ‘greedy hill-climbing’. The above-described search algorithms are designed to explore search spaces systematically; as a result, these approaches only address a single category of problems: observable, deterministic, known environments where the solution is a sequence of actions. However, there are alternative formulation: for different task settings. If the path to the goal does not matter and only attaining the final state is desired, local search algorithms can be considered instead of systematic ones (Russel and Norvig, 2010, p. 120). Local search operates using a limited number of nodes and generally moves only to neighbors of those nodes. Although it is not systematic, local search has two key advantages: (1) it uses very little memory, and (2) it can often find reasonable solutions in large or infinite (continuous) state spaces for which systematic algorithms are unsuitable (Russel and Norvig, 2010, p. 121). ‘Hill- climbing’ (or ‘greedy local search’, Figure 2.10), ‘simulated annealing’, and ‘evolutionary algorithms’ (Figure 2.11) are treated under this title.
FIGURE 2.10 Local search by ‘greedy hill-climbing’. The above-described search algorithms are designed to explore search spaces systematically; as a result, these approaches only address a single category of problems: observable, deterministic, known environments where the solution is a sequence of actions. However, there are alternative formulation: for different task settings. If the path to the goal does not matter and only attaining the final state is desired, local search algorithms can be considered instead of systematic ones (Russel and Norvig, 2010, p. 120). Local search operates using a limited number of nodes and generally moves only to neighbors of those nodes. Although it is not systematic, local search has two key advantages: (1) it uses very little memory, and (2) it can often find reasonable solutions in large or infinite (continuous) state spaces for which systematic algorithms are unsuitable (Russel and Norvig, 2010, p. 121). ‘Hill- climbing’ (or ‘greedy local search’, Figure 2.10), ‘simulated annealing’, and ‘evolutionary algorithms’ (Figure 2.11) are treated under this title.
FIGURE 2.11 Evolutionary parallel search. The character of the task environment affects the problem-solving approach adopted to handle a problem. For example, the above-described approaches assume that the environment is fully observable and deterministic; therefore, the problems are tackled by offline algorithms (Figure 2.12). Starting with the knowledge of an initial state, the agent can calculate exactly which state results from any sequence of actions and knows which state it is in (Russel and Norvig, 2010, p. 133). There is no need for percepts that would provide new information after each action.
FIGURE 2.11 Evolutionary parallel search. The character of the task environment affects the problem-solving approach adopted to handle a problem. For example, the above-described approaches assume that the environment is fully observable and deterministic; therefore, the problems are tackled by offline algorithms (Figure 2.12). Starting with the knowledge of an initial state, the agent can calculate exactly which state results from any sequence of actions and knows which state it is in (Russel and Norvig, 2010, p. 133). There is no need for percepts that would provide new information after each action.
Online search is helpful in dynamic and nondeterministic domains and is necessary for unknown environments. In a state of ignorance, the agent faces an exploration problem where it must use its actions as experiments to inquire into the situation, in order to be able to make decisions. Design situations strongly exhibit these characteristics and it can be claimed that design requires an on-line, exploratory attitude. Thus, following the problem-solving approach, design processes have been likened to ‘exploratory search’. According to this approach, design is a purposeful and conscious specification of the actions that must be taken at the present to attain some desired future conditions, where alternative courses of action are hypothesized and their effects are predicted and evaluated against the predefined set of desired conditions (Kalay, 1992). If the evaluation finds that the predicted effects of the actions match the desired conditions, the hypotheses are validated. On contrary situations, the actions must be modified to achieve the effects of the proposed actions. This is done through iterative modification of the proposed actions and the desired conditions until a match is achieved. In this approach, the design process can be said to comprise three major tasks (Kalay, 1992):
Online search is helpful in dynamic and nondeterministic domains and is necessary for unknown environments. In a state of ignorance, the agent faces an exploration problem where it must use its actions as experiments to inquire into the situation, in order to be able to make decisions. Design situations strongly exhibit these characteristics and it can be claimed that design requires an on-line, exploratory attitude. Thus, following the problem-solving approach, design processes have been likened to ‘exploratory search’. According to this approach, design is a purposeful and conscious specification of the actions that must be taken at the present to attain some desired future conditions, where alternative courses of action are hypothesized and their effects are predicted and evaluated against the predefined set of desired conditions (Kalay, 1992). If the evaluation finds that the predicted effects of the actions match the desired conditions, the hypotheses are validated. On contrary situations, the actions must be modified to achieve the effects of the proposed actions. This is done through iterative modification of the proposed actions and the desired conditions until a match is achieved. In this approach, the design process can be said to comprise three major tasks (Kalay, 1992):
FIGURE 2.14 External, interpreted, and expected function- behavior-structure representation of design process [Gero and Kannengeisser, 2007]. Because the task obviously requires a high level of intelligence, a straightforward proposal is to incorporate human agents, which would assume the role of the online intelligent agent. During her exploration within a design situation, whenever needs and possibilities overlap, a human agent may assign a series of tasks to her tools or assistants. This conception is in line with the traditional view of how a human agent and a computationa system should collaborate. The human agent would employ all her intuitive and holistic thinking and the computer would carry out repetitive and mechanical tasks in which it excels. It appears plausible to interpret current CAD tools in accordance with this pi However, when the human agents are so involved within this process, it is not initial this description, whether computational techniques would be helpful in this p on the type of task and the specific design process’. Moreover, an approach th into a scheme jus practice. If itis no cture. ly obvious solely by rocess; this would depend at would try to fit design t because the computational system requires it that way would most probably fail in t able to carry out autonomous tasks, a computational system should either enhance the capabilities of the human designer or it should enhance the ease and pleasure o f the design process.
FIGURE 2.14 External, interpreted, and expected function- behavior-structure representation of design process [Gero and Kannengeisser, 2007]. Because the task obviously requires a high level of intelligence, a straightforward proposal is to incorporate human agents, which would assume the role of the online intelligent agent. During her exploration within a design situation, whenever needs and possibilities overlap, a human agent may assign a series of tasks to her tools or assistants. This conception is in line with the traditional view of how a human agent and a computationa system should collaborate. The human agent would employ all her intuitive and holistic thinking and the computer would carry out repetitive and mechanical tasks in which it excels. It appears plausible to interpret current CAD tools in accordance with this pi However, when the human agents are so involved within this process, it is not initial this description, whether computational techniques would be helpful in this p on the type of task and the specific design process’. Moreover, an approach th into a scheme jus practice. If itis no cture. ly obvious solely by rocess; this would depend at would try to fit design t because the computational system requires it that way would most probably fail in t able to carry out autonomous tasks, a computational system should either enhance the capabilities of the human designer or it should enhance the ease and pleasure o f the design process.
TABLE 2.1 Ontological categories in situated FBS ontology. Locating function, behavior, and structure of a process in each of these worlds results in nine ontological categories, which is given in Table 2.1 (Gero and Kannengeisser, 2007).
TABLE 2.1 Ontological categories in situated FBS ontology. Locating function, behavior, and structure of a process in each of these worlds results in nine ontological categories, which is given in Table 2.1 (Gero and Kannengeisser, 2007).
TABLE 2.2 Mapping of fundamental steps and activities in situated FBS ontology Additionally, the framework represents external requirements related to the function (FR°), behavior (BR®), and structure (SR°) of processes. Finally, in Figure 2.14, the 20 activities that connect these categories extend the ontology to explore the notion of situated processes (Gero and Kannengeisser, 2007). Gero (1990) proposed eight fundamental steps in design, which were mapped by Gero and Kannengeisser (2004) onto the 20 activities in the situated FBS ontology (Table 2.2, Gero and Kannengeisser, 2007). The numbers of activities correspond to the activities illustrated in Figure 2.14
TABLE 2.2 Mapping of fundamental steps and activities in situated FBS ontology Additionally, the framework represents external requirements related to the function (FR°), behavior (BR®), and structure (SR°) of processes. Finally, in Figure 2.14, the 20 activities that connect these categories extend the ontology to explore the notion of situated processes (Gero and Kannengeisser, 2007). Gero (1990) proposed eight fundamental steps in design, which were mapped by Gero and Kannengeisser (2004) onto the 20 activities in the situated FBS ontology (Table 2.2, Gero and Kannengeisser, 2007). The numbers of activities correspond to the activities illustrated in Figure 2.14
FIGURE 2.15 Basic constituents of a design situation. Yet, there is another kind of complexity, which concerns the negotiated aspects of design. At any design situation there will always be competing viewpoints, which have to be settled through negotiation. This issue was amongst the reasons why design has been characterized as “wicked” (Rittel and Webber, 1973; Buchanan, 1992). This type of complexity is related to the essentially undefined aspects of design, because it only appears within a particular situation and it can only be solved by dispute or negotiation. Therefore, this property of design situations will be referred to as the “inherent undefinedness” of design. A diagram of design situations is given in Figure 2.15, which maps the basic issues that contribute to both types of complexity. Throughout a design process, components of the co-evolving problems and solutions generate an inherent complexity, while a multitude of agents continuously bring forward differing viewpoints and framings, exacerbating the inherent undefinedness.
FIGURE 2.15 Basic constituents of a design situation. Yet, there is another kind of complexity, which concerns the negotiated aspects of design. At any design situation there will always be competing viewpoints, which have to be settled through negotiation. This issue was amongst the reasons why design has been characterized as “wicked” (Rittel and Webber, 1973; Buchanan, 1992). This type of complexity is related to the essentially undefined aspects of design, because it only appears within a particular situation and it can only be solved by dispute or negotiation. Therefore, this property of design situations will be referred to as the “inherent undefinedness” of design. A diagram of design situations is given in Figure 2.15, which maps the basic issues that contribute to both types of complexity. Throughout a design process, components of the co-evolving problems and solutions generate an inherent complexity, while a multitude of agents continuously bring forward differing viewpoints and framings, exacerbating the inherent undefinedness.
FIGURE 2.16 Asummary of the basic problems of design situations (left side) and the strategies and skills of human designers (right side). The diagram in Figure 2.16 brings together and summarizes the basic issues that have been discussed until this point. The column on the left side summarizes the main problems that human or artificial design agents have to face within design situations. Asummary of the strategies and capabilities that human designers utilize for tackling these problems are given on the right side. Design situations are both complex and vague. Human designers employ both everyday intelligence and expert skills to dynamically (re)frame the situations, strategically focus on subsets of tasks, and propose tentative solutions, throughout a co-evolution of problems and solutions. The skills summarized on the right side are the very skills that artificial design agents would require to carry out design.
FIGURE 2.16 Asummary of the basic problems of design situations (left side) and the strategies and skills of human designers (right side). The diagram in Figure 2.16 brings together and summarizes the basic issues that have been discussed until this point. The column on the left side summarizes the main problems that human or artificial design agents have to face within design situations. Asummary of the strategies and capabilities that human designers utilize for tackling these problems are given on the right side. Design situations are both complex and vague. Human designers employ both everyday intelligence and expert skills to dynamically (re)frame the situations, strategically focus on subsets of tasks, and propose tentative solutions, throughout a co-evolution of problems and solutions. The skills summarized on the right side are the very skills that artificial design agents would require to carry out design.
FIGURE 2.17 Basic constituents of the design_proxy approach.
FIGURE 2.17 Basic constituents of the design_proxy approach.
FIGURE 3.1 Satellite dish holder boom; left, initial regular design; right, geometry of the final optimized design [Keane and Brown, 1996) Evolutionary Computation approaches have already been tested in a diverse array of areas. This diverse array includes several design and art related fields as well. Indeed, EC has seen a wide variety of applications in arts and design fields. In design fields, a straightforward idea has been to use Genetic Algorithms (GAs) to optimize parametrically defined two or three-dimensional shapes. In these cases, the genotype most often consists of the parameters that govern a shape of a product and the evaluation is carried out over the phenotype representation. A seminal example of this approach is the evolutionary design of a boom that would hold a satellite dish (Keane and Brown, 1996). The strange shape of the resulting design, while better than the initial regular shape in terms of structural vibration control, was by no means expectable (Figure 3.1). EC has been very popular in generative art circles; at least since, in his book “The Blind Watchmaker”, Richard Dawkins described how evolution could be used to evolve shapes (Dawkins, 1996, pp. 43-74 Lewis, 2008). Following Dawkins, and the Genetic Programming (GP) variant, Karl Sims and William
FIGURE 3.1 Satellite dish holder boom; left, initial regular design; right, geometry of the final optimized design [Keane and Brown, 1996) Evolutionary Computation approaches have already been tested in a diverse array of areas. This diverse array includes several design and art related fields as well. Indeed, EC has seen a wide variety of applications in arts and design fields. In design fields, a straightforward idea has been to use Genetic Algorithms (GAs) to optimize parametrically defined two or three-dimensional shapes. In these cases, the genotype most often consists of the parameters that govern a shape of a product and the evaluation is carried out over the phenotype representation. A seminal example of this approach is the evolutionary design of a boom that would hold a satellite dish (Keane and Brown, 1996). The strange shape of the resulting design, while better than the initial regular shape in terms of structural vibration control, was by no means expectable (Figure 3.1). EC has been very popular in generative art circles; at least since, in his book “The Blind Watchmaker”, Richard Dawkins described how evolution could be used to evolve shapes (Dawkins, 1996, pp. 43-74 Lewis, 2008). Following Dawkins, and the Genetic Programming (GP) variant, Karl Sims and William
FIGURE 3.2 An exploratory task structure for evolutionary design. In order to conceptualize the cooperation of humans, EC-based systems, and additional technologies, a generic task structure is given in Figure 3.2. The task structure develops the exploratory co-evolution model of Maher, Poon, and Boulanger (1996) in two basic ways: First, rather than assuming creativity, emergent behavior, or exploration on behalf of complex EAs, it locates complex EAs only one level above the simple ones. Secondly, in addition to EC systems, separate intelligent agents are posited as necessary, and the exploratory character is assumed to appear only at this level—after the addition of the intelligent agents. In its totality, the task structure in Figure 3.2 may be interpreted as illustrating the required capabilities for a single autonomous evolutionary design agent. However, it is also possible to dismantle this structure towards collaborative models for practical collaboration of human agents and EC based soft-computing systems.
FIGURE 3.2 An exploratory task structure for evolutionary design. In order to conceptualize the cooperation of humans, EC-based systems, and additional technologies, a generic task structure is given in Figure 3.2. The task structure develops the exploratory co-evolution model of Maher, Poon, and Boulanger (1996) in two basic ways: First, rather than assuming creativity, emergent behavior, or exploration on behalf of complex EAs, it locates complex EAs only one level above the simple ones. Secondly, in addition to EC systems, separate intelligent agents are posited as necessary, and the exploratory character is assumed to appear only at this level—after the addition of the intelligent agents. In its totality, the task structure in Figure 3.2 may be interpreted as illustrating the required capabilities for a single autonomous evolutionary design agent. However, it is also possible to dismantle this structure towards collaborative models for practical collaboration of human agents and EC based soft-computing systems.
FIGURE 3.3 Problem structuring (decomposition and integration) through productions.
FIGURE 3.3 Problem structuring (decomposition and integration) through productions.
There are several levels where adaptivity can be implemented. First, adaptivity may concern the whole process and use feedback from the health of the process. Secondly, to each candidate that is being evolved, its own operator parameters can be assigned. Thirdly, each component of each individual may be assigned its own parameters. Therefore, it can be claimed that in adaptive evolution, dynamic characters are assigned to each process, each candidate, or each component. In the last two cases, the feedback procedure is implicit. The underlying idea in these cases is that the operators and parameter settings that have created a better individual should be the better operators and settings. Therefore, although the overall mechanisms would still be fixed, the reproductive operators (crossover and mutation) can be co-evolved together with the products.
There are several levels where adaptivity can be implemented. First, adaptivity may concern the whole process and use feedback from the health of the process. Secondly, to each candidate that is being evolved, its own operator parameters can be assigned. Thirdly, each component of each individual may be assigned its own parameters. Therefore, it can be claimed that in adaptive evolution, dynamic characters are assigned to each process, each candidate, or each component. In the last two cases, the feedback procedure is implicit. The underlying idea in these cases is that the operators and parameter settings that have created a better individual should be the better operators and settings. Therefore, although the overall mechanisms would still be fixed, the reproductive operators (crossover and mutation) can be co-evolved together with the products.
FIGURE 3.6 Schematic illustration of an Interleaved EA process. FIGURE 3.7 Basic structure of the Interleaved EA.
FIGURE 3.6 Schematic illustration of an Interleaved EA process. FIGURE 3.7 Basic structure of the Interleaved EA.
FIGURE 3.8 Potential parallelization of design_proxy-based assistants.
FIGURE 3.8 Potential parallelization of design_proxy-based assistants.
FIGURE 4.1 Graphic arrangement task. The graphic task for developing and demonstrating Interleaved EA and d_p.graphics is defined as generating graphical arrangements using a canvas, basic design units (DUs, i.e., pattern stamps and text strings), and color sources (radial gradients and plain colors) (Figure 4.1). The aim is to generate graphical arrangements that resemble a series of target images, according to different characteristics—in this case, color distribution and graphic layout.
FIGURE 4.1 Graphic arrangement task. The graphic task for developing and demonstrating Interleaved EA and d_p.graphics is defined as generating graphical arrangements using a canvas, basic design units (DUs, i.e., pattern stamps and text strings), and color sources (radial gradients and plain colors) (Figure 4.1). The aim is to generate graphical arrangements that resemble a series of target images, according to different characteristics—in this case, color distribution and graphic layout.
FIGURE 4.2 Genotype for d_p.graphics. In the self-adaptive version, for each objective function, separate lists of standard deviations for ‘nudge’ and ‘color’ mutations, as well as ‘swap rate’ values are also stored and evolved as follows:
FIGURE 4.2 Genotype for d_p.graphics. In the self-adaptive version, for each objective function, separate lists of standard deviations for ‘nudge’ and ‘color’ mutations, as well as ‘swap rate’ values are also stored and evolved as follows:
FIGURE 4.6 Cell, row, and column preparation for Cell and Sequence objectives.
FIGURE 4.6 Cell, row, and column preparation for Cell and Sequence objectives.
FIGURE 4.5 Penalty calculations for Out + Overlap objective.
FIGURE 4.5 Penalty calculations for Out + Overlap objective.
There are a series of process parameters, which should be initiated before evolution starts**. Care should be taken for the adjustment of these parameters, because success of the process is closely connected with proper settings (Figure 4.8).
There are a series of process parameters, which should be initiated before evolution starts**. Care should be taken for the adjustment of these parameters, because success of the process is closely connected with proper settings (Figure 4.8).
FIGURE 4.8 Basic parameters of the d_p.graphics application.
FIGURE 4.8 Basic parameters of the d_p.graphics application.
FIGURE 4.10 Adaptive determination of the current objective.
FIGURE 4.10 Adaptive determination of the current objective.
FIGURE 4.11 Modifications in case of stagnation (for d_p.graphics implementation).
FIGURE 4.11 Modifications in case of stagnation (for d_p.graphics implementation).
FIGURE 4.12 Single objective test series through the desktop wallpaper scenario; 10 runs of 200 generations for each objective. Non-adaptive and process-adaptive versions are given, both use the best-so-far initial parameters.
FIGURE 4.12 Single objective test series through the desktop wallpaper scenario; 10 runs of 200 generations for each objective. Non-adaptive and process-adaptive versions are given, both use the best-so-far initial parameters.
FIGURE 4.13 Comparison of non-adaptive and self-adaptive versions with best parameter settings. Single-objective runs for Cell objective. Each graph is the average of 10 runs.
FIGURE 4.13 Comparison of non-adaptive and self-adaptive versions with best parameter settings. Single-objective runs for Cell objective. Each graph is the average of 10 runs.
FIGURE 4.14 Comparison of non-adaptive and self-adaptive versions with poor parameter settings. Single-objective runs for Cell objective. Each graph is the average of 10 runs.
FIGURE 4.14 Comparison of non-adaptive and self-adaptive versions with poor parameter settings. Single-objective runs for Cell objective. Each graph is the average of 10 runs.
FIGURE 4.15 Inputs and results of the first Cofee-cup scenario, regular and abundant (upper right) cases.
FIGURE 4.15 Inputs and results of the first Cofee-cup scenario, regular and abundant (upper right) cases.
FIGURE 4.16 Examples of multi-objective runs: Coffee-cup 2 scenario and its abundant version (process-adaptive). Layout objectives converged rather fast, as a result of initiation through initiation maps. After this fast convergence, while the Color objective continued to work for improving the color distribution, layout objectives simply maintained their fitness levels, which is exactly the aim of the naive Interleaved EA. It should be noted that the Out + Overlap objective was set to rather tolerant overlap values for these scenarios, which required a degree of overlap. These trials show that the stochastic nature of the EAs, when coupled with tolerant objective functions, exhibit a flexible design approach. In other words, it is possible to produce a variety of similar, yet distinct solutions for a graphic design task; solutions, which comprise different types and numbers of DUs, while still resembling the same targets.
FIGURE 4.16 Examples of multi-objective runs: Coffee-cup 2 scenario and its abundant version (process-adaptive). Layout objectives converged rather fast, as a result of initiation through initiation maps. After this fast convergence, while the Color objective continued to work for improving the color distribution, layout objectives simply maintained their fitness levels, which is exactly the aim of the naive Interleaved EA. It should be noted that the Out + Overlap objective was set to rather tolerant overlap values for these scenarios, which required a degree of overlap. These trials show that the stochastic nature of the EAs, when coupled with tolerant objective functions, exhibit a flexible design approach. In other words, it is possible to produce a variety of similar, yet distinct solutions for a graphic design task; solutions, which comprise different types and numbers of DUs, while still resembling the same targets.
FIGURE 4.17 Results of the Desktop 1 scenario, non-abundant versions (process-adaptive).
FIGURE 4.17 Results of the Desktop 1 scenario, non-abundant versions (process-adaptive).
FIGURE 4.18 Results of the Desktop 2 scenario, abundant and non-abundant versions together (process-adaptive).
FIGURE 4.18 Results of the Desktop 2 scenario, abundant and non-abundant versions together (process-adaptive).
FIGURE 4.19 Results of the T-shirt 1 scenario, abundant and non-abundant versions together (process-adaptive).
FIGURE 4.19 Results of the T-shirt 1 scenario, abundant and non-abundant versions together (process-adaptive).
FIGURE 4.20 Results of the T-shirt 2 scenario, abundant and non-abundant versions together (process-adaptive).
FIGURE 4.20 Results of the T-shirt 2 scenario, abundant and non-abundant versions together (process-adaptive).
FIGURE 4.22 Example processes and results for the Pareto-based version (left) and self-adaptive naive Interleaved EA (rightmost column)
FIGURE 4.22 Example processes and results for the Pareto-based version (left) and self-adaptive naive Interleaved EA (rightmost column)
FIGURE 4.23 The effect of the initiation maps.
FIGURE 4.23 The effect of the initiation maps.
This model can be seen as a kind of integral decomposition approach. As can be seen in Figure 4.24, there may be distinct tasks within an evolutionary design process. These tasks may concern, for example, production of a background image, production of the schematic layout, insertion of the texts, etc. If we descend one level down, within each task layer, problem-setting procedures and evolutionary process may be separated. At this level, the evolutionary procedure may be a complex EA, such as the Interleaved EA. When we descend one more level, each of the procedures for setting the parameters of the process may be separated from the others. At this level, a multitude of agents may collaborate with each other for the definition of separate aspects of the process. Furthermore, as mentioned above, this phase can be considered as an independent and continuous process. A user team may define a specific problem, and at the mean time may benefit from ready-made specifications of an EA expert. They can also use existing style-sets for their aims.
This model can be seen as a kind of integral decomposition approach. As can be seen in Figure 4.24, there may be distinct tasks within an evolutionary design process. These tasks may concern, for example, production of a background image, production of the schematic layout, insertion of the texts, etc. If we descend one level down, within each task layer, problem-setting procedures and evolutionary process may be separated. At this level, the evolutionary procedure may be a complex EA, such as the Interleaved EA. When we descend one more level, each of the procedures for setting the parameters of the process may be separated from the others. At this level, a multitude of agents may collaborate with each other for the definition of separate aspects of the process. Furthermore, as mentioned above, this phase can be considered as an independent and continuous process. A user team may define a specific problem, and at the mean time may benefit from ready-made specifications of an EA expert. They can also use existing style-sets for their aims.
FIGURE 4.25 Levels for a hierarchical decomposition of architecture.
FIGURE 4.25 Levels for a hierarchical decomposition of architecture.
FIGURE 4.26 A layer-based task decomposition approach. The layer based decomposition approach has been loosely based on the horizontal decomposition scheme of Rodney Brooks (Figure 4.27). Ina seminal study, Brooks (1999, originally published in 1986) proposed to convert from the temporal hierarchical decomposition of the then dominant Al approaches to a parallel yet still hierarchical architecture. Instead of breaking the process into a chain of information processing modules as in Figure 4.27a, Brooks proposed a decomposition in terms of behavior generating modules, each of which connects sensing to action (Figure 4.27b). The scheme is still hierarchical in the sense that layers are added incrementally, and newer layers depend on earlier layers (Brooks, 1999). While brooks’ decomposition model is rather concrete with an aimed robotic structure, ours is deliberately kept abstract. Although our model will also be elaborated into a decomposition and integration scheme for architecture, it will not be carried over to an architectural application with multiple layers. The application will integrate several layers into a single problem definition, which will not be claimed to test for the real significance of the approach. This necessitated the open-ended description of the idea on a rather overall level, with an attempt to set forth a new research framework while situating our applications over this framework.
FIGURE 4.26 A layer-based task decomposition approach. The layer based decomposition approach has been loosely based on the horizontal decomposition scheme of Rodney Brooks (Figure 4.27). Ina seminal study, Brooks (1999, originally published in 1986) proposed to convert from the temporal hierarchical decomposition of the then dominant Al approaches to a parallel yet still hierarchical architecture. Instead of breaking the process into a chain of information processing modules as in Figure 4.27a, Brooks proposed a decomposition in terms of behavior generating modules, each of which connects sensing to action (Figure 4.27b). The scheme is still hierarchical in the sense that layers are added incrementally, and newer layers depend on earlier layers (Brooks, 1999). While brooks’ decomposition model is rather concrete with an aimed robotic structure, ours is deliberately kept abstract. Although our model will also be elaborated into a decomposition and integration scheme for architecture, it will not be carried over to an architectural application with multiple layers. The application will integrate several layers into a single problem definition, which will not be claimed to test for the real significance of the approach. This necessitated the open-ended description of the idea on a rather overall level, with an attempt to set forth a new research framework while situating our applications over this framework.
FIGURE 4.27 (a) Atraditional decomposition of a mobile robot control system into functional modules, (b) A decomposition of a mobile robot control system based on task achieving behaviors [Brooks, 1999].
FIGURE 4.27 (a) Atraditional decomposition of a mobile robot control system into functional modules, (b) A decomposition of a mobile robot control system based on task achieving behaviors [Brooks, 1999].
FIGURE 4.28 A layer based task decomposition scheme for conceptual architectural design.
FIGURE 4.28 A layer based task decomposition scheme for conceptual architectural design.
FIGURE 4.29 An abstract 3D formation as an architectura stem cell.
FIGURE 4.29 An abstract 3D formation as an architectura stem cell.
FIGURE 4.30 Three example ASCs, manually developed within the task-based decomposition framework.
FIGURE 4.30 Three example ASCs, manually developed within the task-based decomposition framework.
FIGURE 4.31 Example target, ground floor of Santa Monica Library by Moore Ruble Yudell Architects. Vector painted on Inkscape according to given layers and swatches (seen on the right). An example ‘painted’ target for d_p.layout is given in Figure 4.31. On the right side a collection of vector painting styles are given. Each style corresponds to a separate layer. Using these layers, the system parses the ‘svg’ document and defines each shape as ‘something’, such as a zone, a plan boundary, a front, an open facade, a kind of functional DU, a scale line, north arrow, a road, etc. Note that until this point the approach remains within the limits of atomism. The scene comprises discrete items with definite properties. The context of the architectural layout problem allows this kind of discretization, to the extent that it corresponds to the set of items that is being used by human designers.
FIGURE 4.31 Example target, ground floor of Santa Monica Library by Moore Ruble Yudell Architects. Vector painted on Inkscape according to given layers and swatches (seen on the right). An example ‘painted’ target for d_p.layout is given in Figure 4.31. On the right side a collection of vector painting styles are given. Each style corresponds to a separate layer. Using these layers, the system parses the ‘svg’ document and defines each shape as ‘something’, such as a zone, a plan boundary, a front, an open facade, a kind of functional DU, a scale line, north arrow, a road, etc. Note that until this point the approach remains within the limits of atomism. The scene comprises discrete items with definite properties. The context of the architectural layout problem allows this kind of discretization, to the extent that it corresponds to the set of items that is being used by human designers.
FIGURE 4.32 Example target simplified (only DUs), ground floor of Santa Monica Library.
FIGURE 4.32 Example target simplified (only DUs), ground floor of Santa Monica Library.
FIGURE 4.33 Library functions. In each column, painting layers (left side) is mapped to a simplified library program (right side)
FIGURE 4.33 Library functions. In each column, painting layers (left side) is mapped to a simplified library program (right side)
Although d_p.layout can start its process with a predefined and fixed set of DUs, it can also assume the task to generate this list of DUs by itself, from an analysis of a pool of targets, in which case the task definition becomes, properly determining and arranging a list of DUs over a series of building floors with variable plan boundaries (Figure 4.35). d_p.layout carries out this task with regard to a series of example layouts, and in its current applications it is dependent on well-established typologies like libraries, schools, and museums. If this last point is also included in the task definition, we can state that the task of d_p.layout is properly determining and arranging a list of DUs with regard to a series of example layouts, over specific floors of a building that is within a specific typology. In its most basic expression, the task of the d_p.layout is defined as adequately populating a series of plan borders, i.e., several floors of a building with a fixed set of arbitrary polygonal DUs, whose forms and dimensions are also determined and fixed at the outset (Figure 4.34). Here, the term ‘adequately’ refers to various considerations in terms of functionalities, performances, formal qualities, and constraints. Therefore, the basic goal is a proper placement and arrangement of DUs with regard to each other and to the given contextual elements and conditions.
Although d_p.layout can start its process with a predefined and fixed set of DUs, it can also assume the task to generate this list of DUs by itself, from an analysis of a pool of targets, in which case the task definition becomes, properly determining and arranging a list of DUs over a series of building floors with variable plan boundaries (Figure 4.35). d_p.layout carries out this task with regard to a series of example layouts, and in its current applications it is dependent on well-established typologies like libraries, schools, and museums. If this last point is also included in the task definition, we can state that the task of d_p.layout is properly determining and arranging a list of DUs with regard to a series of example layouts, over specific floors of a building that is within a specific typology. In its most basic expression, the task of the d_p.layout is defined as adequately populating a series of plan borders, i.e., several floors of a building with a fixed set of arbitrary polygonal DUs, whose forms and dimensions are also determined and fixed at the outset (Figure 4.34). Here, the term ‘adequately’ refers to various considerations in terms of functionalities, performances, formal qualities, and constraints. Therefore, the basic goal is a proper placement and arrangement of DUs with regard to each other and to the given contextual elements and conditions.
FIGURE 4.35 Layout generation for multiple floors with arbitrary outlines.
FIGURE 4.35 Layout generation for multiple floors with arbitrary outlines.
FIGURE 4.36 Candidate preparation and initiation. Left: painted candidate, middle: initiated candidate, right: best result of evolution. The initial placement of DUs is part of candidate initiation. Fixed DUs are initiated within their pre- specified position. There is an additional option to initiate a series of movable DUs within specific locations to be modified during evolution. The remaining DUs are initiated within the bounding box of a given floor. This bounding box is divided into a grid and each DU is randomly assigned to a grid point (Figure 4.36). The need for such grid initiation is linked with the integrated evolution of circulation elements. These elements are represented simply as another group of DUs. Therefore, to facilitate the intermingling of regular DUs and circulation elements, an optional ‘striped initiation’ is developed. Circulation elements are initiated in their own rows and columns, which are regularly interleaved to the other rows and columns. A candidate preparation snapshot is given in Figure 4.36. This candidate is a simplified test version of one of the used targets (Halmstad Library by Schmidt Hammer Lassen Architects). The painted candidate can be seen on the left. Note that movable DUs are later taken out of the borders of the candidate. The middle image is an initiated candidate with the painted DUs. Note that the central green area is a fixed element.
FIGURE 4.36 Candidate preparation and initiation. Left: painted candidate, middle: initiated candidate, right: best result of evolution. The initial placement of DUs is part of candidate initiation. Fixed DUs are initiated within their pre- specified position. There is an additional option to initiate a series of movable DUs within specific locations to be modified during evolution. The remaining DUs are initiated within the bounding box of a given floor. This bounding box is divided into a grid and each DU is randomly assigned to a grid point (Figure 4.36). The need for such grid initiation is linked with the integrated evolution of circulation elements. These elements are represented simply as another group of DUs. Therefore, to facilitate the intermingling of regular DUs and circulation elements, an optional ‘striped initiation’ is developed. Circulation elements are initiated in their own rows and columns, which are regularly interleaved to the other rows and columns. A candidate preparation snapshot is given in Figure 4.36. This candidate is a simplified test version of one of the used targets (Halmstad Library by Schmidt Hammer Lassen Architects). The painted candidate can be seen on the left. Note that movable DUs are later taken out of the borders of the candidate. The middle image is an initiated candidate with the painted DUs. Note that the central green area is a fixed element.
FIGURE 4.39 Modification of the objective priority list for d_p.layout. Whenever a new individual is created within the process, it has to be evaluated for all the objectives of that evolutionary run, in order to be able to find the ranks of the candidates. However, time expenditure of each objective function varies considerably. For example, the Sequence objective consumes more time than the others. Therefore, it was practically desirable to diminish the total number of mutation numbers for such expensive objectives. In the Interleaved EA, each objective can be assigned a different mutation number, and the current objective’s mutation number would be used at a specific generation. However, this was not solving the above-mentioned problem, because after each mutation, newly created individuals still had to be evaluated in terms of all objectives, including the slower ones, forcing the user to adjust to the slowest objective. Therefore, a new option is added to the system. After the specified number of mutations are carried out, newly created individuals are first evaluated only in terms of the current objective. Then their number is diminished to a pre-specified amount according to only the current objective and through the selection procedure of the current objective. Only after this reduction, the candidate population is evaluated for the remaining objectives. This procedure enables the specification of higher mutation numbers for faster objectives; hence searching through more neighboring nodes becomes possible at a specific state of the population at least for faster objectives.
FIGURE 4.39 Modification of the objective priority list for d_p.layout. Whenever a new individual is created within the process, it has to be evaluated for all the objectives of that evolutionary run, in order to be able to find the ranks of the candidates. However, time expenditure of each objective function varies considerably. For example, the Sequence objective consumes more time than the others. Therefore, it was practically desirable to diminish the total number of mutation numbers for such expensive objectives. In the Interleaved EA, each objective can be assigned a different mutation number, and the current objective’s mutation number would be used at a specific generation. However, this was not solving the above-mentioned problem, because after each mutation, newly created individuals still had to be evaluated in terms of all objectives, including the slower ones, forcing the user to adjust to the slowest objective. Therefore, a new option is added to the system. After the specified number of mutations are carried out, newly created individuals are first evaluated only in terms of the current objective. Then their number is diminished to a pre-specified amount according to only the current objective and through the selection procedure of the current objective. Only after this reduction, the candidate population is evaluated for the remaining objectives. This procedure enables the specification of higher mutation numbers for faster objectives; hence searching through more neighboring nodes becomes possible at a specific state of the population at least for faster objectives.
As in the previous application, the aim of the Out + Overlap objective is to keep DUs within plan limits while preventing DU overlap (Figure 4.40). Because the floor boundaries are irregular in the d_p.layout application, more complicated procedures became required for calculating the penalties for the DUs that exceed the floor boundary. Otherwise, the behavior of this procedure is basically the same: give penalty to floor boundary violation and to inner DU border overlaps.
As in the previous application, the aim of the Out + Overlap objective is to keep DUs within plan limits while preventing DU overlap (Figure 4.40). Because the floor boundaries are irregular in the d_p.layout application, more complicated procedures became required for calculating the penalties for the DUs that exceed the floor boundary. Otherwise, the behavior of this procedure is basically the same: give penalty to floor boundary violation and to inner DU border overlaps.
The Trivial Hole objective is much simpler (Figure 4.42). It simply calculates the ratio of the occupied area within plan borders, so that maximizing this value amounts to the minimizing of DU overlap and outer boundary violation; and implicitly, minimizing the empty areas within boundaries. However, there is no explicit mechanism for eliminating trivial holes between DUs. Thus, perhaps Proximity fitness would better be named Trivial Hole in terms of its functioning. For the Proximity objective, the DUs are checked for collisions of their inner and outer borders (Figure 4.41). If two DUs collide only on their outer borders, without colliding on the inner, an award is given that is proportional with the legitimate overlap area. Thus, this objective works for keeping the DUs close and aligned. The offset value for the outer border controls the proximity of the DUs. A similar procedure checks for the DUs’ neighborhood with the outer borders. To find the Proximity fitness of a candidate layout, outer boundary violation penalties are subtracted from proximity awards.
The Trivial Hole objective is much simpler (Figure 4.42). It simply calculates the ratio of the occupied area within plan borders, so that maximizing this value amounts to the minimizing of DU overlap and outer boundary violation; and implicitly, minimizing the empty areas within boundaries. However, there is no explicit mechanism for eliminating trivial holes between DUs. Thus, perhaps Proximity fitness would better be named Trivial Hole in terms of its functioning. For the Proximity objective, the DUs are checked for collisions of their inner and outer borders (Figure 4.41). If two DUs collide only on their outer borders, without colliding on the inner, an award is given that is proportional with the legitimate overlap area. Thus, this objective works for keeping the DUs close and aligned. The offset value for the outer border controls the proximity of the DUs. A similar procedure checks for the DUs’ neighborhood with the outer borders. To find the Proximity fitness of a candidate layout, outer boundary violation penalties are subtracted from proximity awards.
The above-mentioned formal objectives depend on the inner characteristics of candidate layouts. Functional objectives, on the other hand, are measured with reference to target buildings. The Sequence objective (Figure 4.43) extracts sequences from target buildings and tries to make new plans that include those sequences. In the d_p.graphics version, the absolute positions of the DUs with regard to the canvas were being considered. However, in the plan generation example, the important information is the relative positioning of the functional units (DUs) with regard to each other. Therefore, the task is reformulated as a sequence matching procedure. As such, it isa more developed version, yet still following the methods of Smith and Chang (1999).
The above-mentioned formal objectives depend on the inner characteristics of candidate layouts. Functional objectives, on the other hand, are measured with reference to target buildings. The Sequence objective (Figure 4.43) extracts sequences from target buildings and tries to make new plans that include those sequences. In the d_p.graphics version, the absolute positions of the DUs with regard to the canvas were being considered. However, in the plan generation example, the important information is the relative positioning of the functional units (DUs) with regard to each other. Therefore, the task is reformulated as a sequence matching procedure. As such, it isa more developed version, yet still following the methods of Smith and Chang (1999).
FIGURE 4.43 Calculation of the Sequence fitness.
FIGURE 4.43 Calculation of the Sequence fitness.
For the Neighbor objective, a target is prepared as follows: The neighborhoods of example buildings are painted with a simple diagram. The direction that a DU is neighboring is also given with this diagram. Target preparation amounts to collecting the frequency of neighborhoods between DU types within a neighborhood matrix, from all floors of a single example building. The rows and columns of the matrix are symmetrical and their sequence is fixed for all applications. The frequencies of neighborhood types are normalized by the maximum frequency. This way the matrix becomes a neighborhood pattern, which can be compared with other buildings.
For the Neighbor objective, a target is prepared as follows: The neighborhoods of example buildings are painted with a simple diagram. The direction that a DU is neighboring is also given with this diagram. Target preparation amounts to collecting the frequency of neighborhoods between DU types within a neighborhood matrix, from all floors of a single example building. The rows and columns of the matrix are symmetrical and their sequence is fixed for all applications. The frequencies of neighborhood types are normalized by the maximum frequency. This way the matrix becomes a neighborhood pattern, which can be compared with other buildings.
For candidate evaluation, the target's matrix is compared with the candidate’s. As with the Neighbor objective, Neighbor Cell objective calculates and negates the difference between two matrix patterns while the Adjacency Cell version calculates the negative of the total penalty. Finally, there is a group of grid cell objectives. Similar to the Neighbor objective, the Neighbor Cell objective (Figure 4.45) tries to evolve plans that are similar to a target neighbor matrix distribution, according to DU type. However, in this case the analysis method for the spatial adjacencies is different. In addition, adjacencies to directions are not taken into account for the Cell type objectives. When used together the two objective types enable a more fine grained comparison with a target library. Also separate targets may be used for each objective. For fitness calculation, a fixed sized square query cell is placed over each DU’s center. Types of all the DUs that collide with a DU's query cell are added to the cell neighbor list of that DU. The query cell is square instead of circular, because this way the overlaps can be directly determined through R-tree indexing, which makes the analysis procedure very efficient, albeit a bit coarser; because in this approach DU collisions are calculated with regard to DUs’ bounding boxes.
For candidate evaluation, the target's matrix is compared with the candidate’s. As with the Neighbor objective, Neighbor Cell objective calculates and negates the difference between two matrix patterns while the Adjacency Cell version calculates the negative of the total penalty. Finally, there is a group of grid cell objectives. Similar to the Neighbor objective, the Neighbor Cell objective (Figure 4.45) tries to evolve plans that are similar to a target neighbor matrix distribution, according to DU type. However, in this case the analysis method for the spatial adjacencies is different. In addition, adjacencies to directions are not taken into account for the Cell type objectives. When used together the two objective types enable a more fine grained comparison with a target library. Also separate targets may be used for each objective. For fitness calculation, a fixed sized square query cell is placed over each DU’s center. Types of all the DUs that collide with a DU's query cell are added to the cell neighbor list of that DU. The query cell is square instead of circular, because this way the overlaps can be directly determined through R-tree indexing, which makes the analysis procedure very efficient, albeit a bit coarser; because in this approach DU collisions are calculated with regard to DUs’ bounding boxes.
FIGURE 4.46 Calculation of the Manual Cell fitness. Finally, the Manual Cell objective (Figure 4.46) tries to bring several DU types together by looking for cells that conform to manually given cell lists. This objective is developed for obtaining already known functional groupings, like building cores, which frequently bring toilets, elevators, staircases, and technical elements together. To achieve its aim, the procedure regularly divides the bounding box of a candidate layout and traverses through each grid point with a square query cell for sampling the existing groupings. Each sample is compared with given desired cell groupings. The aim is to find all the DU types close together as overlapping a query cell. Each query cell returns a ratio within range 0-1] that reveals its success. A ratio of "1" denotes that all desired DU types are found together within a cell. Whenever a value of “1” is found, the procedure ends. Otherwise, it keeps searching through the whole grid and returns the best-found ratio. If there are more than one desired cell lists, the average value of all the best-found ratios is returned.
FIGURE 4.46 Calculation of the Manual Cell fitness. Finally, the Manual Cell objective (Figure 4.46) tries to bring several DU types together by looking for cells that conform to manually given cell lists. This objective is developed for obtaining already known functional groupings, like building cores, which frequently bring toilets, elevators, staircases, and technical elements together. To achieve its aim, the procedure regularly divides the bounding box of a candidate layout and traverses through each grid point with a square query cell for sampling the existing groupings. Each sample is compared with given desired cell groupings. The aim is to find all the DU types close together as overlapping a query cell. Each query cell returns a ratio within range 0-1] that reveals its success. A ratio of "1" denotes that all desired DU types are found together within a cell. Whenever a value of “1” is found, the procedure ends. Otherwise, it keeps searching through the whole grid and returns the best-found ratio. If there are more than one desired cell lists, the average value of all the best-found ratios is returned.
FIGURE 4.47 Trivial Hole objective, for 50 DUs. ii) Trivial Hole objective, best results of four example runs, 50 DUs, Halmstad Library scenario
FIGURE 4.47 Trivial Hole objective, for 50 DUs. ii) Trivial Hole objective, best results of four example runs, 50 DUs, Halmstad Library scenario
i) Candidate preparation and initiation, 50 DUs, Halmstad Library scenario
i) Candidate preparation and initiation, 50 DUs, Halmstad Library scenario
i) Candidate preparation and initiation, 150 DUs, Halmstad Library scenario
i) Candidate preparation and initiation, 150 DUs, Halmstad Library scenario
FIGURE 4.48 Trivial Hole objective, for 150 DUs. ii) Trivial Hole objective, best results of four example runs, 150 DUs, Halmstad Library scenario
FIGURE 4.48 Trivial Hole objective, for 150 DUs. ii) Trivial Hole objective, best results of four example runs, 150 DUs, Halmstad Library scenario
ii) Trivial Hole objective, best results of four example runs, 250 DUs, Halmstad Library scenario
ii) Trivial Hole objective, best results of four example runs, 250 DUs, Halmstad Library scenario
FIGURE 4.49 Trivial Hole objective, for 250 DUs. Visual inspection of the results in Figure 4.49 shows that the amount of unwanted overlaps have increased in this scenario (the rightmost example gives only inner borders), which is not surprising. However, the main problem for such a high DU number is rather a potential problem with overall organization. Indeed, even the most complicated example layouts did not exceed a number of 150 DUs throughout our experimentations. And apparently, handling such a complicated plan in only one stage is not a viable method. If such large and complicated plans are first separated into higher regions corresponding to main functional groupings, then these could be handled separately, which would also result in more manageable problems; hence a hierarchical approach appears more appropriate for these situations. Still, although not perfect, it appears that the system is able to function for the Trivial Hole objective even with such a high DU number.
FIGURE 4.49 Trivial Hole objective, for 250 DUs. Visual inspection of the results in Figure 4.49 shows that the amount of unwanted overlaps have increased in this scenario (the rightmost example gives only inner borders), which is not surprising. However, the main problem for such a high DU number is rather a potential problem with overall organization. Indeed, even the most complicated example layouts did not exceed a number of 150 DUs throughout our experimentations. And apparently, handling such a complicated plan in only one stage is not a viable method. If such large and complicated plans are first separated into higher regions corresponding to main functional groupings, then these could be handled separately, which would also result in more manageable problems; hence a hierarchical approach appears more appropriate for these situations. Still, although not perfect, it appears that the system is able to function for the Trivial Hole objective even with such a high DU number.
FIGURE 4.50 Target and candidate (tight fit) for Sequence objective. Target, candidate, and initiation for the first test series can be seen in Figure 4.50. The candidate is almost the same with the target. In Figure 4.51, results of a series of control tests are given. These examples are evolved only with the Trivial Hole objective, thus their functional arrangements are totally random. These will enable us to assess if the test results could be obtained randomly. For the sake of simplicity, rotation mutations have been disabled for these series. The situation is more complicated for the functional objectives. The main requirement is to produce targe direc t-candidate pairs that would reveal whether an objective guides an evolution in the desired tion or not. As has been stated above, each functional objective is paired with a formal one for these tests. The most complicated verification problem is raised by the Sequence objective. The basic question is, “will the sequences given within the target layout reappear in the candidates?” The fitnes: function is so designed that the sequences do not require an absolute matching of the target and candi date layout arrangements. It just tries to maximize the most frequent sequence types in the candi date. The tests will start with the simplest case, which comes close to searching for an absolute similarity of the target and candidate. Then through a more complicated case, the real functioning of theo bjective will be examined.
FIGURE 4.50 Target and candidate (tight fit) for Sequence objective. Target, candidate, and initiation for the first test series can be seen in Figure 4.50. The candidate is almost the same with the target. In Figure 4.51, results of a series of control tests are given. These examples are evolved only with the Trivial Hole objective, thus their functional arrangements are totally random. These will enable us to assess if the test results could be obtained randomly. For the sake of simplicity, rotation mutations have been disabled for these series. The situation is more complicated for the functional objectives. The main requirement is to produce targe direc t-candidate pairs that would reveal whether an objective guides an evolution in the desired tion or not. As has been stated above, each functional objective is paired with a formal one for these tests. The most complicated verification problem is raised by the Sequence objective. The basic question is, “will the sequences given within the target layout reappear in the candidates?” The fitnes: function is so designed that the sequences do not require an absolute matching of the target and candi date layout arrangements. It just tries to maximize the most frequent sequence types in the candi date. The tests will start with the simplest case, which comes close to searching for an absolute similarity of the target and candidate. Then through a more complicated case, the real functioning of theo bjective will be examined.
FIGURE 4.51 Control evolution with only Trivial Hole objective (tight fit scenario). The best results of two test series (population: 150, mutation: 15, crossover: QO) are given in Figure 4.52. The first row concerns the simplest case, which involves short sequences of length 4 and 5 that have been searched only within rows and columns of the candidates (without 45° rotation). The tests in the second row take both 0° and 45° rotated row and columns into account. Each image is the best Sequence objective result of a single evolutionary run. On the left, 10000 generation tests are given, while 20000 generation tests are given on the right side.
FIGURE 4.51 Control evolution with only Trivial Hole objective (tight fit scenario). The best results of two test series (population: 150, mutation: 15, crossover: QO) are given in Figure 4.52. The first row concerns the simplest case, which involves short sequences of length 4 and 5 that have been searched only within rows and columns of the candidates (without 45° rotation). The tests in the second row take both 0° and 45° rotated row and columns into account. Each image is the best Sequence objective result of a single evolutionary run. On the left, 10000 generation tests are given, while 20000 generation tests are given on the right side.
FIGURE 4.52 Sequence objective verification - Sequence + Trivial Hole, tight fit candidate, best results of O° and 0° + 45° versions - Sequence grid width: 2 meters.
FIGURE 4.52 Sequence objective verification - Sequence + Trivial Hole, tight fit candidate, best results of O° and 0° + 45° versions - Sequence grid width: 2 meters.
FIGURE 4.53 Sequence + Trivial Hole, example fitness graphs.
FIGURE 4.53 Sequence + Trivial Hole, example fitness graphs.
FIGURE 4.54 Sequence objective verification - Sequence and Out + Overlap, loose fit candidate, best results of O° and 0° + 45° versions — Sequence grid width: 2 m. Asecond series of similar tests with the Out + Overlap objective can be seen in Figure 4.54. With slight differences in the problem setting, the results are similar. Example fitness graphs are given in Figure 4.55.
FIGURE 4.54 Sequence objective verification - Sequence and Out + Overlap, loose fit candidate, best results of O° and 0° + 45° versions — Sequence grid width: 2 m. Asecond series of similar tests with the Out + Overlap objective can be seen in Figure 4.54. With slight differences in the problem setting, the results are similar. Example fitness graphs are given in Figure 4.55.
FIGURE 4.55 Sequence and Out + Overlap, example fitness graphs.
FIGURE 4.55 Sequence and Out + Overlap, example fitness graphs.
FIGURE 4.56 Target, candidate, and initiation for Sequence objective.
FIGURE 4.56 Target, candidate, and initiation for Sequence objective.
Generation: 9969; T.h. Fitness: 71.750 Generation: 9826; T.h. Fitness: 71.835 FIGURE 4.57 Random control runs with only Trivial Hole objective.
Generation: 9969; T.h. Fitness: 71.750 Generation: 9826; T.h. Fitness: 71.835 FIGURE 4.57 Random control runs with only Trivial Hole objective.
Generation: 19878; O.o. Fitness: -10.570; Seq. Fitness: 8.020 Generation: 19964; O.o. Fitness: -9.108; Seq. Fitness: 11.114
Generation: 19878; O.o. Fitness: -10.570; Seq. Fitness: 8.020 Generation: 19964; O.o. Fitness: -9.108; Seq. Fitness: 11.114
FIGURE 4.59 Sequence and Out + Overlap, example fitness graphs — sequences [3, 4, 5], rotations [0°, 45°]. FIGURE 4.58 Sequence objective verification - Sequence and Out + Overlap - Sequence grid width: 2 m.
FIGURE 4.59 Sequence and Out + Overlap, example fitness graphs — sequences [3, 4, 5], rotations [0°, 45°]. FIGURE 4.58 Sequence objective verification - Sequence and Out + Overlap - Sequence grid width: 2 m.
Generation: 19902; O.o. Fitness: -9.813; Seq. Fitness: 20.544
Generation: 19902; O.o. Fitness: -9.813; Seq. Fitness: 20.544
Generation: 19947; T.h. Fitness: 74.909; Adj. Fitness: -100.287 Generation: 19938; T.h. Fitness: 74.211; Adj. Fitness: -101.260 Generation: 19925; T.h. Fitness: 74.046; Neigh. Fitness: -110.867 Generation: 19998; T.h. Fitness: 74.165; Neigh. Fitness: -107.24
Generation: 19947; T.h. Fitness: 74.909; Adj. Fitness: -100.287 Generation: 19938; T.h. Fitness: 74.211; Adj. Fitness: -101.260 Generation: 19925; T.h. Fitness: 74.046; Neigh. Fitness: -110.867 Generation: 19998; T.h. Fitness: 74.165; Neigh. Fitness: -107.24
FIGURE 4.60 Neighbor and Adjacency objectives; target, candidate, and initiation.
FIGURE 4.60 Neighbor and Adjacency objectives; target, candidate, and initiation.
FIGURE 4.62 Neighbor + Trivial Hole, example fitness graphs. The fitness graphs in Figure 4.62 and 4.63 indicate that the processes had not yet converged in 20000 generations. However, in practice, there will rarely be a longer period for an objective and these rather intermediate results are sufficient for the evaluation of the tendencies of these fitness functions. The main difference between the two objective functions is that the Adjacency fitness can be maximized with over-exploiting the most advantageous DU relationships; light blue and light green in this scenario. However, the Neighbor objective penalizes a disruption of the target pattern, hence it becomes difficult to lean too much on the exploitation of any relationship on its own; the overall neighborhood pattern has to be conserved. This can be evaluated by the inspection of yellow and dark blue cells. Because their overall fitness gain is lower, the Adjacency objective tends to omit improving the neighborhood patterns of these cell types and concentrates on the more frequent DU types, so that we can observe free yellow and dark blue cells with no desired neighborhood at all. On the other hand, as expected, the Neighbor objective appears more consistent in improving with all types of DUs and mostly succeeds in regenerating the target neighborhood patterns.
FIGURE 4.62 Neighbor + Trivial Hole, example fitness graphs. The fitness graphs in Figure 4.62 and 4.63 indicate that the processes had not yet converged in 20000 generations. However, in practice, there will rarely be a longer period for an objective and these rather intermediate results are sufficient for the evaluation of the tendencies of these fitness functions. The main difference between the two objective functions is that the Adjacency fitness can be maximized with over-exploiting the most advantageous DU relationships; light blue and light green in this scenario. However, the Neighbor objective penalizes a disruption of the target pattern, hence it becomes difficult to lean too much on the exploitation of any relationship on its own; the overall neighborhood pattern has to be conserved. This can be evaluated by the inspection of yellow and dark blue cells. Because their overall fitness gain is lower, the Adjacency objective tends to omit improving the neighborhood patterns of these cell types and concentrates on the more frequent DU types, so that we can observe free yellow and dark blue cells with no desired neighborhood at all. On the other hand, as expected, the Neighbor objective appears more consistent in improving with all types of DUs and mostly succeeds in regenerating the target neighborhood patterns.
FIGURE 4.63 Adjacency + Trivial Hole, example fitness graphs. Asimilar distinction applies to the Neighbor Cell and its Adjacency Cell version. The targets and candidates for these objectives are given in Figure 4.64. In this setting, for the Neighbor Cell objective, a dark blue DU would like to be near two light blue DUs and a red DU. Likewise, a light blue DU would prefer the neighborhood of one light blue, one red, and one dark blue DUs. In the Adjacency Cell version, because both dark blue and red DUs award light blue neighbors more than the others, the light blue DUs would tend to get collected near each other. However, the same cause would affect
FIGURE 4.63 Adjacency + Trivial Hole, example fitness graphs. Asimilar distinction applies to the Neighbor Cell and its Adjacency Cell version. The targets and candidates for these objectives are given in Figure 4.64. In this setting, for the Neighbor Cell objective, a dark blue DU would like to be near two light blue DUs and a red DU. Likewise, a light blue DU would prefer the neighborhood of one light blue, one red, and one dark blue DUs. In the Adjacency Cell version, because both dark blue and red DUs award light blue neighbors more than the others, the light blue DUs would tend to get collected near each other. However, the same cause would affect
Generation: 9951; T.H. fitness: 76.930; N. Cell fitness: -134.734 Generation: 19720; T.H. fitness: 77.873; A. Cell fitness: -97.329
Generation: 9951; T.H. fitness: 76.930; N. Cell fitness: -134.734 Generation: 19720; T.H. fitness: 77.873; A. Cell fitness: -97.329
Generation: 19720; T.H. fitness: 78.069; N. Cell fitness: -128.451 FIGURE 4.64 Neighbor Cell and Adjacency Cell objectives; target, candidate, and initiation.
Generation: 19720; T.H. fitness: 78.069; N. Cell fitness: -128.451 FIGURE 4.64 Neighbor Cell and Adjacency Cell objectives; target, candidate, and initiation.
Generation: 19984; T.H. fitness: 77.779; A. Cell fitness: -97.295
Generation: 19984; T.H. fitness: 77.779; A. Cell fitness: -97.295
FIGURE 4.67 Adjacency Cell + Trivial Hole, example fitness graphs. FIGURE 4.66 Neighbor Cell + Trivial Hole, example fitness graphs. Finally, the test setting and example results for the Manual Cell objective can be seen in Figure 4.68 and 4.69. The test tasks are rather easy and as can be seen in Figure 4.70, in an incremental progression (which indicates the attaining of desired proximities one by one), the process converges in just a few hundred generations. Although this test setting is simple (indeed the initiation example in Figure 4.71 has already found the desired cells right through random initiation, which is not usually the case), even the hard tasks for the Manual Cell objective would remain relatively simple compared to the other objectives, and the desired fitness states are generally reached very quickly.
FIGURE 4.67 Adjacency Cell + Trivial Hole, example fitness graphs. FIGURE 4.66 Neighbor Cell + Trivial Hole, example fitness graphs. Finally, the test setting and example results for the Manual Cell objective can be seen in Figure 4.68 and 4.69. The test tasks are rather easy and as can be seen in Figure 4.70, in an incremental progression (which indicates the attaining of desired proximities one by one), the process converges in just a few hundred generations. Although this test setting is simple (indeed the initiation example in Figure 4.71 has already found the desired cells right through random initiation, which is not usually the case), even the hard tasks for the Manual Cell objective would remain relatively simple compared to the other objectives, and the desired fitness states are generally reached very quickly.
FIGURE 4.70 Manual Cell + Trivial Hole, example fitness graphs.
FIGURE 4.70 Manual Cell + Trivial Hole, example fitness graphs.
FIGURE 4.69 Manual Cell objective, example best results - Grid width: 4m, query cell width: 8m.
FIGURE 4.69 Manual Cell objective, example best results - Grid width: 4m, query cell width: 8m.
i) Manual Cell fitness, target 1
i) Manual Cell fitness, target 1
Fitness graphs (for mean fitness) - average of 100 trials - 6 objectives - 50 DUs
Fitness graphs (for mean fitness) - average of 100 trials - 6 objectives - 50 DUs
FIGURE 4.72 Prepared outlines and fixed DUs of a five-floor example run.
FIGURE 4.72 Prepared outlines and fixed DUs of a five-floor example run.
FIGURE 4.73 Fitness graphs for the ground floor of the example run.
FIGURE 4.73 Fitness graphs for the ground floor of the example run.
FIGURE 4.75 Workshop process for Evolutionary Collectivity Izmir workshop.
FIGURE 4.75 Workshop process for Evolutionary Collectivity Izmir workshop.
FIGURE 4.76 Images from the Evolutionary Collectivity Izmir workshop, from brainstorming sessions to final presentations. The design process started with a collective brainstorming session (see Figure 4.76 for workshop photographs). Through several stages, from concepts expressed with a few words, the students moved forward to sketched out and diagrammed spatial ideas. The process started with the participation of the whole group. Further on, the participants were divided into groups of eight, and eventually they continued the workshop individually or as teams of two to three people. At this stage, they were asked to choose their favorite spatial ideas and start working on the main problem of the workshop; that is, designing a medium-scale library within the given site. The site for Izmir workshop was right across Yasar University Campus, and was easily accessible. It was almost flat, in between a metro line and a highway, next to a metro station and was being crossed by a pedestrian bridge, so that there were many situational parameters in need of reconciliation.
FIGURE 4.76 Images from the Evolutionary Collectivity Izmir workshop, from brainstorming sessions to final presentations. The design process started with a collective brainstorming session (see Figure 4.76 for workshop photographs). Through several stages, from concepts expressed with a few words, the students moved forward to sketched out and diagrammed spatial ideas. The process started with the participation of the whole group. Further on, the participants were divided into groups of eight, and eventually they continued the workshop individually or as teams of two to three people. At this stage, they were asked to choose their favorite spatial ideas and start working on the main problem of the workshop; that is, designing a medium-scale library within the given site. The site for Izmir workshop was right across Yasar University Campus, and was easily accessible. It was almost flat, in between a metro line and a highway, next to a metro station and was being crossed by a pedestrian bridge, so that there were many situational parameters in need of reconciliation.
FIGURE 4.78 Muge and Ceren - left: initial sketches, middle: first model, right: sketch submitted for d_p.layout. The first example is the team of Mige Halici and Ceren Abacioglu, who were 4" year students at the time of the workshop. They started their examination with the keywords, “non-accessible”, “spaciousness”, and “airy”, which they developed into a conception that understands interior space of a building in terms of “largeness and extensiveness” (Figure 4.78). They then moved forward to question the relationship between “inside” and “outside” and set forth a hierarchical spatial organization from “protected” to “unprotected”. At this point, they envisaged their building as a cube that is pierced by the pedestrian bridge that crosses over the site. During the night, the d_p.layout processed seven floors of this large cube. It turned out that the team made a mistake about the height of the bridge; however, this was no problem at all. The proposal consisted of just floor outlines, and it was the d_p.layout who worked through the night. In the first results for the seven floors (Figure 4.79), the layouts include no vertical circulation elements or distinct circulation areas. Functional zones are not coordinated between floors, yet they involve a sufficient variety of functional elements that are arranged with regard to their functionalities. It is difficult for a human designer to delve into such detail at this stage, because she has to think about a large group of contextual and conceptual factors before moving into detailed arrangements. However, in this case there is almost no cost of bringing out these draft arrangements.
FIGURE 4.78 Muge and Ceren - left: initial sketches, middle: first model, right: sketch submitted for d_p.layout. The first example is the team of Mige Halici and Ceren Abacioglu, who were 4" year students at the time of the workshop. They started their examination with the keywords, “non-accessible”, “spaciousness”, and “airy”, which they developed into a conception that understands interior space of a building in terms of “largeness and extensiveness” (Figure 4.78). They then moved forward to question the relationship between “inside” and “outside” and set forth a hierarchical spatial organization from “protected” to “unprotected”. At this point, they envisaged their building as a cube that is pierced by the pedestrian bridge that crosses over the site. During the night, the d_p.layout processed seven floors of this large cube. It turned out that the team made a mistake about the height of the bridge; however, this was no problem at all. The proposal consisted of just floor outlines, and it was the d_p.layout who worked through the night. In the first results for the seven floors (Figure 4.79), the layouts include no vertical circulation elements or distinct circulation areas. Functional zones are not coordinated between floors, yet they involve a sufficient variety of functional elements that are arranged with regard to their functionalities. It is difficult for a human designer to delve into such detail at this stage, because she has to think about a large group of contextual and conceptual factors before moving into detailed arrangements. However, in this case there is almost no cost of bringing out these draft arrangements.
FIGURE 4.79 Muge and Ceren, results of the first trial (seven floors, max. gen.: 7000, pop.: 150, cross.: 0, mut.: 15, Manual Cell target ((“vc", “t", “t”, “t" "we", “we")), Neighbor Target = Neighbor Cell target: Santa Monica, Sequence target: Free University Berlin, using max. %92 of the available plan areas, DU offset: 1.2m, Sequence grid width: 1.5m, M. Cell grid width: 8m, query cell widths: 16m, DU area limit: 45m2, DU width limit: 8m, formal refinement on with mut.:75, diminish: 25) It can be noticed in Figure 4.79 that the initial DU rotation functionality failed to approximate layout orientations and the DUs are collectively placed in a diagonal orientation. The system fitted lines using the four corner points of the floor outlines, which would be different with another orientation. Whenever a high number of DUs are wrongly orientated, they have to adapt to the floor outlines in a collective manner, because their orientation is also connected to the neighboring DUs. This is a complicated combinatorial task that would most probably be omitted, while the system can maximize its fitness through easier operations towards sub-optimal areas*®.
FIGURE 4.79 Muge and Ceren, results of the first trial (seven floors, max. gen.: 7000, pop.: 150, cross.: 0, mut.: 15, Manual Cell target ((“vc", “t", “t”, “t" "we", “we")), Neighbor Target = Neighbor Cell target: Santa Monica, Sequence target: Free University Berlin, using max. %92 of the available plan areas, DU offset: 1.2m, Sequence grid width: 1.5m, M. Cell grid width: 8m, query cell widths: 16m, DU area limit: 45m2, DU width limit: 8m, formal refinement on with mut.:75, diminish: 25) It can be noticed in Figure 4.79 that the initial DU rotation functionality failed to approximate layout orientations and the DUs are collectively placed in a diagonal orientation. The system fitted lines using the four corner points of the floor outlines, which would be different with another orientation. Whenever a high number of DUs are wrongly orientated, they have to adapt to the floor outlines in a collective manner, because their orientation is also connected to the neighboring DUs. This is a complicated combinatorial task that would most probably be omitted, while the system can maximize its fitness through easier operations towards sub-optimal areas*®.
FIGURE 4.80 Muge and Ceren, solid-void exercise over d_p.layout drafts.
FIGURE 4.80 Muge and Ceren, solid-void exercise over d_p.layout drafts.
FIGURE 4.81 MUge and Ceren, sketches submitted for second d_p.layout iteration.
FIGURE 4.81 MUge and Ceren, sketches submitted for second d_p.layout iteration.
FIGURE 4.82 Muge and Ceren, results of the second iteration (six floors, max. gen.: 8000, DU area limit: 30m2, DU width limit: 4.5m)
FIGURE 4.82 Muge and Ceren, results of the second iteration (six floors, max. gen.: 8000, DU area limit: 30m2, DU width limit: 4.5m)
The second example from this workshop is the case of Idil Kantarci, which was special. Idil was a first year student and had never attended to an architectural design studio. She was a basic design student with absolutely no experience in architectural design because of her school’s curriculum. Nevertheless, she was very eager and serious.
The second example from this workshop is the case of Idil Kantarci, which was special. Idil was a first year student and had never attended to an architectural design studio. She was a basic design student with absolutely no experience in architectural design because of her school’s curriculum. Nevertheless, she was very eager and serious.
FIGURE 4.84 Idil Kantarci's first d_p.layout trial, as prepared in Inkscape by the tutors.
FIGURE 4.84 Idil Kantarci's first d_p.layout trial, as prepared in Inkscape by the tutors.
FIGURE 4.85 Idil's first d_p.layout results (DU area limit: 48m2, DU width limit: 8m) When Idil brought her first results the next morning (Figure 4.85), it was seen that the last floor was not finished (the draft in the figure developed during the day). A reason was a failure in DU width and area limit settings by the tutor, which could have been far larger with such large floors. Nevertheless, detailed programming and patching studies had been obtained for three floors, from out of almost nothing. Examining these results, it appeared that Idil’s total area was far larger than was required, which was far easier to discuss over these detailed scale drafts (consider discussing such issues with a novice that has no understanding of scale).
FIGURE 4.85 Idil's first d_p.layout results (DU area limit: 48m2, DU width limit: 8m) When Idil brought her first results the next morning (Figure 4.85), it was seen that the last floor was not finished (the draft in the figure developed during the day). A reason was a failure in DU width and area limit settings by the tutor, which could have been far larger with such large floors. Nevertheless, detailed programming and patching studies had been obtained for three floors, from out of almost nothing. Examining these results, it appeared that Idil’s total area was far larger than was required, which was far easier to discuss over these detailed scale drafts (consider discussing such issues with a novice that has no understanding of scale).
FIGURE 4.86 Left: Idil's solid-void exercise for two floors, right: preparation for the second trial. After obtaining her first results, Idil became able to work on her solid - void model with reference to the draft layouts (Figure 4.86). She had never made building models. This can be observed from her flat spatial models. Nevertheless, through her studies on these drafts, she started to see primary elements of a library layout, circulation system, articulation of spaces, the types of DUs, and so on. Examining d_p.layout outputs amounted to a crash-course on layout generation for Idil and this initiation proved enough to render her productive. She quickly moved towards a better understanding of the spatial organization and started shaping it with her sketches. At the same time, she started to understand how to use d_p.layout, and she started to refine and rationalize her drafts in an attempt to gain control of the process.
FIGURE 4.86 Left: Idil's solid-void exercise for two floors, right: preparation for the second trial. After obtaining her first results, Idil became able to work on her solid - void model with reference to the draft layouts (Figure 4.86). She had never made building models. This can be observed from her flat spatial models. Nevertheless, through her studies on these drafts, she started to see primary elements of a library layout, circulation system, articulation of spaces, the types of DUs, and so on. Examining d_p.layout outputs amounted to a crash-course on layout generation for Idil and this initiation proved enough to render her productive. She quickly moved towards a better understanding of the spatial organization and started shaping it with her sketches. At the same time, she started to understand how to use d_p.layout, and she started to refine and rationalize her drafts in an attempt to gain control of the process.
FIGURE 4.87 Idil's second trial. Top: initial submission printed after being prepared and parsed. Bottom: outputs of the trial in their original layouts (DU area limit: 50m2, DU width limit: 8m).
FIGURE 4.87 Idil's second trial. Top: initial submission printed after being prepared and parsed. Bottom: outputs of the trial in their original layouts (DU area limit: 50m2, DU width limit: 8m).
FIGURE 4.89 Idil Kantarci's final product (top-left: ground floor, top-right: first floor over the ground floor). Idil's final output was quite surprising in its refined articulation of functional areas and intricate, rich, and well-developed interior spaces (Figure 4.89). She collected the multimedia areas on the northern side of the building within a two storey high L-shaped space where a forest of spherical cellular elements would be hanged down to be seen from both the highway and from inside the building through transparent dividers. The central courtyard, which was shaped through a playful cooperation with d_p.layout, presents a sensitive scale and a pleasant spatial articulation. Commercial area, restaurant, offices, activity hall, library areas, and services are all placed and shaped sensitively and in a rational manner. A fascinating result from a first year student's very first venture into architecture. Indeed, better than most of the fourth year student teams.
FIGURE 4.89 Idil Kantarci's final product (top-left: ground floor, top-right: first floor over the ground floor). Idil's final output was quite surprising in its refined articulation of functional areas and intricate, rich, and well-developed interior spaces (Figure 4.89). She collected the multimedia areas on the northern side of the building within a two storey high L-shaped space where a forest of spherical cellular elements would be hanged down to be seen from both the highway and from inside the building through transparent dividers. The central courtyard, which was shaped through a playful cooperation with d_p.layout, presents a sensitive scale and a pleasant spatial articulation. Commercial area, restaurant, offices, activity hall, library areas, and services are all placed and shaped sensitively and in a rational manner. A fascinating result from a first year student's very first venture into architecture. Indeed, better than most of the fourth year student teams.
FIGURE 4.88 Idil's sketches from the third day.
FIGURE 4.88 Idil's sketches from the third day.
FIGURE 4.90 Photographs from the Evolutionary Collectivity Mardin workshop.
FIGURE 4.90 Photographs from the Evolutionary Collectivity Mardin workshop.
FIGURE 4.91 Workshop process for the Evolutionary Collectivity Mardin workshop. Busra Tekelioglu and Merve Cap started their work with the concept of “relaxedness”, and tried to reconcile their ideas with contextual conditions in their first model (Figure 4.92), which was dissolved into four masses to adapt to the scale of the scattered neighborhood, while at the same tin creating inner streets. The d_p.layout settings and parameters were mostly the same, with the important exception of the usage of only four objectives (Trivial Hole, Sequence, Neighbor Cell, and Manual Cell). This meant better average results for each of these four objectives with longer evolution due to slow convergence, which was mostly permitted. Therefore, some floors of the trials could not be finished on time, i.e., in one night. In addition, because the interface was not fail-safe, problems arose in one team’s trials, which could not be compensated within the fast progression of the workshop and this team was forced to work over premature results despite the team members’ eagerness. Two example processes that have interacted freely and productively with the d_p.layout will be depicted in the following pages.
FIGURE 4.91 Workshop process for the Evolutionary Collectivity Mardin workshop. Busra Tekelioglu and Merve Cap started their work with the concept of “relaxedness”, and tried to reconcile their ideas with contextual conditions in their first model (Figure 4.92), which was dissolved into four masses to adapt to the scale of the scattered neighborhood, while at the same tin creating inner streets. The d_p.layout settings and parameters were mostly the same, with the important exception of the usage of only four objectives (Trivial Hole, Sequence, Neighbor Cell, and Manual Cell). This meant better average results for each of these four objectives with longer evolution due to slow convergence, which was mostly permitted. Therefore, some floors of the trials could not be finished on time, i.e., in one night. In addition, because the interface was not fail-safe, problems arose in one team’s trials, which could not be compensated within the fast progression of the workshop and this team was forced to work over premature results despite the team members’ eagerness. Two example processes that have interacted freely and productively with the d_p.layout will be depicted in the following pages.
FIGURE 4.92 Busra Tekelioglu and Merve Cap, top-left: initial massing, right: ground floor of d_p.layout output (DU area limit: 130m2, DU width limit: 10m), bottom-left: solid-void models.
FIGURE 4.92 Busra Tekelioglu and Merve Cap, top-left: initial massing, right: ground floor of d_p.layout output (DU area limit: 130m2, DU width limit: 10m), bottom-left: solid-void models.
FIGURE 4.93 Second day sketches of Bisra and Merve, from d_p.layout drafts, toward final models. Sequence articulations and Neighbor Cell groupings can be clearly seen in Figure 4.92 The solid-void studies gave way to the idea of an inner gorge traversing along the length of the masses, which was reflected in the final proposal. The team’s progression towards final plans and models can be examined in Figure 4.93. The left column presents functional zoning studies. On the top right, studies on the inner gorge can be seen. On the lower right, these studies can be seen as integrated into schematic drawings.
FIGURE 4.93 Second day sketches of Bisra and Merve, from d_p.layout drafts, toward final models. Sequence articulations and Neighbor Cell groupings can be clearly seen in Figure 4.92 The solid-void studies gave way to the idea of an inner gorge traversing along the length of the masses, which was reflected in the final proposal. The team’s progression towards final plans and models can be examined in Figure 4.93. The left column presents functional zoning studies. On the top right, studies on the inner gorge can be seen. On the lower right, these studies can be seen as integrated into schematic drawings.
FIGURE 4.94 Busra and Merve, final model (bottom-left: ground floor, bottom-right: first floor over ground floor). The final proposal of Buisra and Merve can be seen in Figure 4.94. Adaptation to the context appears as a dominant shaping factor for this project. However, in the end, variety of the in-between spaces, the non-monotonous, fractured interior circulations, visual connections between floors, and created inner perspectives make this project unexpectedly rich in terms of spatial development. This richness in the spatial formation somewhat complicates the functional organization. First of all, the separation of library functions into different masses was not questioned. But this may be due to the characteristics of the historical settlement where the university is located. In Mardin, most historical buildings are organized around or next to courtyards, and circulation from exterior spaces is common. Secondly, both because the students were in their second years, and the workshop had to move forward too quickly, the students tended to follow d_p.layout's drafts in a less questioning manner than the Izmir workshop. In other words, they trusted d_p.layout and used its drafts, and this was the only viable way to finish the workshop in such a short period. Nevertheless, considering the state of development, scale, and placement of the masses, and the building's situation within its neighborhood, the project appears rather convincing; quite open for further development. It is visually pleasing and if was built, it would add a positive quality to its rather undefined neighborhood.
FIGURE 4.94 Busra and Merve, final model (bottom-left: ground floor, bottom-right: first floor over ground floor). The final proposal of Buisra and Merve can be seen in Figure 4.94. Adaptation to the context appears as a dominant shaping factor for this project. However, in the end, variety of the in-between spaces, the non-monotonous, fractured interior circulations, visual connections between floors, and created inner perspectives make this project unexpectedly rich in terms of spatial development. This richness in the spatial formation somewhat complicates the functional organization. First of all, the separation of library functions into different masses was not questioned. But this may be due to the characteristics of the historical settlement where the university is located. In Mardin, most historical buildings are organized around or next to courtyards, and circulation from exterior spaces is common. Secondly, both because the students were in their second years, and the workshop had to move forward too quickly, the students tended to follow d_p.layout's drafts in a less questioning manner than the Izmir workshop. In other words, they trusted d_p.layout and used its drafts, and this was the only viable way to finish the workshop in such a short period. Nevertheless, considering the state of development, scale, and placement of the masses, and the building's situation within its neighborhood, the project appears rather convincing; quite open for further development. It is visually pleasing and if was built, it would add a positive quality to its rather undefined neighborhood.
FIGURE 4.95 Merve, Bahri, and Melike, conceptual sections. They have prepared their d_p.layout proposal by fixing a hall that would represent the large entrance ramp on both ground and basement floors (Figure 4.96). Resulting d_p.layout drafts appeared mostly applicable and they became instructive for the team while developing their final proposal, which resulted in dynamic and convenient interiors. Again, the effect of the Sequence (Free University Berlin) and Cell targets (Santa Monica Library) can be clearly observed.
FIGURE 4.95 Merve, Bahri, and Melike, conceptual sections. They have prepared their d_p.layout proposal by fixing a hall that would represent the large entrance ramp on both ground and basement floors (Figure 4.96). Resulting d_p.layout drafts appeared mostly applicable and they became instructive for the team while developing their final proposal, which resulted in dynamic and convenient interiors. Again, the effect of the Sequence (Free University Berlin) and Cell targets (Santa Monica Library) can be clearly observed.
FIGURE 4.96 Merve, Bahri, and Melike, d_p.layout preparation and results (DU area limit: 1OOm2, DU width limit: 8m).
FIGURE 4.96 Merve, Bahri, and Melike, d_p.layout preparation and results (DU area limit: 1OOm2, DU width limit: 8m).
FIGURE 4.97 Merve, Bahri, and Melike, solid - void studies and the first envelope proposal.
FIGURE 4.97 Merve, Bahri, and Melike, solid - void studies and the first envelope proposal.
FIGURE 4.98 Merve, Bahri, and Melike, sketches toward the final model.
FIGURE 4.98 Merve, Bahri, and Melike, sketches toward the final model.
Although available time was severely limited, the Mardin workshop was more streamlined and turned out to be more successful than the previous workshop. As the tutors, we have prepared a common site model to be used by all teams and we were able to finish the workshop with an exhibition, yet without a final presentation session. Although the students had to attend to other courses—including their regular design studio courses—they tried to attend to the workshop as much as possible. In the end, the tutors, students, and the staff of the university appeared content with the results. Although through only one trial, the students interacted with the d_p.layout better and were more inclined to exploit its capabilities.
Although available time was severely limited, the Mardin workshop was more streamlined and turned out to be more successful than the previous workshop. As the tutors, we have prepared a common site model to be used by all teams and we were able to finish the workshop with an exhibition, yet without a final presentation session. Although the students had to attend to other courses—including their regular design studio courses—they tried to attend to the workshop as much as possible. In the end, the tutors, students, and the staff of the university appeared content with the results. Although through only one trial, the students interacted with the d_p.layout better and were more inclined to exploit its capabilities.
TABLE 4.1 Evolutionary Collectivity workshop survey. A convenience of the system is the variability of the control exercised by the user. The amount of control ranges from a more autonomous state (only floor outlines given) to a highly controlled state with elaborate definitions comprising pre-given fixed and mobile DUs, intricate floor shapes, landscape elements, separators, courtyards, directions, etc. The students tended to use these contro possibilities after getting used to both the system and the design problems.
TABLE 4.1 Evolutionary Collectivity workshop survey. A convenience of the system is the variability of the control exercised by the user. The amount of control ranges from a more autonomous state (only floor outlines given) to a highly controlled state with elaborate definitions comprising pre-given fixed and mobile DUs, intricate floor shapes, landscape elements, separators, courtyards, directions, etc. The students tended to use these contro possibilities after getting used to both the system and the design problems.
FIGURE 4.100 Survey results (mean) for Izmir and Mardin workshops.
FIGURE 4.100 Survey results (mean) for Izmir and Mardin workshops.
FIGURE 4.101 design_proxy approach and its application for design_proxy. layout. Before demonstrating the d_p.layout system, the chapter located the architectural layout problem within the broader problem of the conceptual design of buildings. This resulted in the proposal of a conceptual framework for artificial architectural design assistants, i.e., the Architectural Stem Cells (ASC) Framework. The ASC Framework proposes a dynamic and multi-layered method for combining | set of design assistants for larger tasks in architectural design.
FIGURE 4.101 design_proxy approach and its application for design_proxy. layout. Before demonstrating the d_p.layout system, the chapter located the architectural layout problem within the broader problem of the conceptual design of buildings. This resulted in the proposal of a conceptual framework for artificial architectural design assistants, i.e., the Architectural Stem Cells (ASC) Framework. The ASC Framework proposes a dynamic and multi-layered method for combining | set of design assistants for larger tasks in architectural design.
The second component of the ASC Framework is a conception for abstract building development specifications, i.e., Architectural Stem Cells (Figure 4.102). The ASCs are required for re-integrating the separated task layers of the architectural problem through solution-based development (Figure 4.30). An ASC can be conceived as a semantically marked geometric structure, which contains the information that specifies the possibilities and constraints for how an abstract building may develop from an undetailed stage to a fully developed building draft.
The second component of the ASC Framework is a conception for abstract building development specifications, i.e., Architectural Stem Cells (Figure 4.102). The ASCs are required for re-integrating the separated task layers of the architectural problem through solution-based development (Figure 4.30). An ASC can be conceived as a semantically marked geometric structure, which contains the information that specifies the possibilities and constraints for how an abstract building may develop from an undetailed stage to a fully developed building draft.
FIGURE 5.3 Library for the Faculty of Philology, Free University of Berlin.
FIGURE 5.3 Library for the Faculty of Philology, Free University of Berlin.
Related topics:
Evolutionary algorithmsDesign TheoryDigital ArchitectureDesign ComputingFacility Layout PlanningComputational DesignComputer Aided Architectural DesignFacility Layout DesignComputer Aided Design (Architecture)Design Automation
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved