Minds and Machines

, Volume 20, Issue 4, pp 533–564

The ERI-Designer: A Computer Model for the Arrangement of Furniture

Authors

    • División de Ciencias de la Comunicación y DiseñoUniversidad Autónoma Metropolitana, Unidad Cuajimalpa
  • Alfredo Aguilar
    • Posgrado en Ciencias de la ComputaciónUniversidad Nacional Autónoma de México
  • Santiago Negrete
    • División de Ciencias de la Comunicación y DiseñoUniversidad Autónoma Metropolitana, Unidad Cuajimalpa
Article

DOI: 10.1007/s11023-010-9208-9

Cite this article as:
Pérez y Pérez, R., Aguilar, A. & Negrete, S. Minds & Machines (2010) 20: 533. doi:10.1007/s11023-010-9208-9
  • 161 Views

Abstract

This paper reports a computer program to generate novel designs for the arrangement of furniture within a simulated room. It is based on the engagement-reflection computer model of the creative processes. During engagement the system generates material in the form of sequences of actions (e.g. change the colours of the walls, include some furniture in the room, modify their colour, and so on) guided by content and knowledge constraints. During reflection, the system evaluates the composition produced so far and, if it is necessary, modifies it. We discuss the implementation of the system and some of its most salient features, especially the use of a computational model for creativity in the terrain of design. We argue that this kind of model opens new possibilities for the simulation of the design processes as well as the development of tools.

Keywords

Computational creativityEngagement-reflectionDesignFurniture arrangement

Introduction

Interior design is a complex task that involves knowledge about environmental psychology, architecture, product design, and even anthropology [see the comments on Whitemyer (2009) about the work of Laurier, e.g. (2008a, b)]. It is out of the scope of this paper, and today probably impossible, to represent the whole interior design process in computer terms. However, using it as a framework, we have developed a system that produces novel and interesting arrangement of furniture in a given room. We introduce the term Computer Interior Design (CID) to make explicit the difference between the complex products that human interior design is capable of producing and the limited interior design outcomes that our computer program can generate. Nevertheless, this type of program provides insights about interior design that otherwise would be difficult to produce.

Computer models of design have taken the attention of many researchers in the past. AI applications have very early defined the process as a problem solving one. Under this banner several techniques have been investigated exploiting search-space traversal methods. Amongst those is the use of expert systems (Gero 1985). For example, Maher (1985) describes an expert system called HI-RISE for structural design for construction. It is based on a frame-based production-system language and LISP.

More recently, sub-symbolic approaches have also been used in design. The approach described in Coyne et al. (1993), serves to create room content descriptions (a list of the elements present in a room regardless of their distribution) using a connectionist style learning algorithm that acts over the links of a graph of room descriptors (furniture or other room features). The links of the graph denote the likeliness of the two nodes being together in the same type of room (bedroom, bathroom, kitchen, etc.). The algorithm weakens or reinforces the links by exposure to a number of examples of successful room designs previously stored in a database. The technique reinforces links of room descriptors that occur more often in the examples and weakens those that occur less often or do not occur at all. The result yields room descriptions that are original but also keep the intuitive relationship between the elements and yet allow for rooms with unusual content to crop up (e.g. kitchen with a TV set). The algorithm to strengthen or weaken the links successfully recovers and combines previous designs to produce new and novel ones. The rationale behind the decisions taken to arrive at such designs remains alien to the model. As it often happens with connectionist methods, it is difficult to explain, in terms of the domain, why the output resulted.

Genetic Algorithms have been used to solve configuration problems in design (Woodbury 1993). However, finding a suitable coding that both facilitates the process and the application of the selection rule and provides a simple intuitive realization mechanism is often difficult. Design tools have been developed in the past years with two main emphases in mind: design laboratories and design assistants. The first group focuses on freeing the designer from burdensome tasks that consume time and offer a workbench where ideas can be quickly prototyped, tested, simulated, etc. so that the best options can be chosen and developed further. CAD systems fall in this category. Design assistants are systems where the focus of the tool is to participate in the design process (e.g. Janssen et al. 2002; Gómez de Silva and Zamora 2004) by generating possible designs that not only bring possible solutions to the designer but can also suggest interesting and novel ideas that had not been considered before in the problem. In this category the systems can or cannot emulate the way humans work but it is often the case that systems are developed with some sort of model of human cognition in mind.

The program that we report in this paper follows a different approach to those described. We envision CID as a composition process, where the elements that comprise the composition satisfy a set of constraints. By contrast with previous works, we believe that interactions between those elements elicit emotional and affective reactions in the designer. This claim seems to be supported by professional interior designers. For example, in the introductory video entitled Perspectives in Interior Design that can be found in the web page of the International Interior Design Association (http://www.iida.org), Mitch Sawasy affirms that: “Designs are emotions. When we walk into a space we have an emotional response”. Researchers in related fields support this idea. For example, Pullman and Gross (2004) study how to create emotional nexus with guests or customers through careful planning of tangible and intangible service elements within a business; what they are interested in is analyzing emotional responses as mediating factors between the physical and relational elements that conform customers’ experience and their loyalty behaviour. Thus, many design decisions are based on emotions (e.g. Lewis and Haviland-Jones 2004). Following Gelernter, who affirms that emotions are the glue of ideas during the creative process (Gelernter 1994, p. 5), we claim that it is possible to use computational representations of affective reactions to the environment as cues to probe memory in order to progress composition during the CID process. As we will show, the use of such affective reactions provides the required flexibility to produce a composition.

To test our claim we employ the engagement-reflection (E-R) computer model of creativity (Pérez y Pérez and Sharples 2001). The main idea behind the model is that the creative process is formed by two main phases: the generation phase, called engagement, where ideas to progress a composition are generated. The typical example of engagement is daydreaming, where sequences of actions to advance a creative work (e.g. a composition, a narrative, and so on) are produced. The second phase is called reflection, where the ideas produced so far are evaluated to check if they satisfy the requirements of the current task. If it is necessary, such ideas are modified. Then, a new engagement phase starts again and the cycle continues. We believe that the interplay between engagement and reflection is an essential force that drives the creative process. Thus, this work concentrates on developing a computer representation of some of the affective reactions elicited during the interior design process, and then implementing the E-R model in a computer program for CID in order to generate novel compositions of furniture in a given room.

In this work, a composition is considered novel when it is not possible to find a similar one in the system’s knowledge-base. We name our program The Engagement-Reflection Interior Designer (ERI-Designer). This is a description of how it works: During engagement the ERI-Designer generates material in the form of sequences of actions (e.g. change the colours of the walls, include some furniture in the room, modify their colour, and so on) guided by content and knowledge constraints. During reflection, the system evaluates the composition produced so far and, if it is necessary, modifies it. As part of the evaluation, the system updates the constraints that drive engagement, influencing in this way the generation phase. Then, the system switches back to engagement and the cycle continues until the composition is finished. In the following lines we describe how we represent affective reactions in our system, the characteristics of the room and household goods we work with, how the system creates its knowledge base, how the system generates a design and how we evaluated it.

Affective Reactions

It is out of the scope of this paper to carry out any study related to individuals’ affective reactions to spatial arrangements of furniture. Therefore, for the present version of the program, and based on the work of experts in the area, we have defined 7 rules, known as CID-Rules, to detect such situations. Thus, if a given composition breaks any of the CID-rules, an affective reaction, referred to as tension, is triggered. The following lines describe each rule:
  1. 1.

    Colour Harmony. Colour harmony depends on the eye of the beholder (Fehrman and Fehrman 2004). Nevertheless, in this work we assume that the use of analogue colours, i.e. adjacent colours within the chromatic circle, produce a pleasant perception. Thus, if a composition involves the use of non-analogue colours, a tension due to colour harmony is triggered.

     
  2. 2.

    Colour Contrasts. Following Mahnke (1996), in interior design contrast establishes how furniture is highlighted or to some extent hidden. However, one must act carefully because “Color contrasts may be helpful or harmful[...] Here are some general hints: 1. Hues similar in saturation and value can unify a room and make a space seem larger. However, be sure to avoid monotony 2. Contrast between walls and furnishings will make the furnishings more prominent[...]” (pp. 85–86). In this work we are interested in finding a nice balance between contrasts. So, if the composition hardly includes contrasts, a tension due to monotony is triggered; in the same way, if the composition includes too much contrast so that “it does more harm than good” to the observer, a tension due to excessive contrast is triggered.

     
  3. 3.

    Colour Impression. Fehrman and Fehrman (2004) describe a study about colour and interior environments made by Masao Inui in Japan: “Since the color impression of a room is primarily experienced as an integrated experience by an observer, Inui found that the less bright the color (heading toward neutrality), the more pleasant the interior was thought to be.” (p. 128). Following this author, in this work if the integrated experience of colour impression trends towards light-neutral colours, the composition is considered as pleasant. Otherwise, a tension due to colour impression is triggered. The elements considered to calculate the colour impression are the four walls of the room and the dominant colours of the furniture. Since the colour of the floor is fixed in this prototype, the current version of the ERI-Designer does not take it into consideration.

     
  4. 4.

    Functional Value. Ritterfeld (2002) points out the importance of the functional value of quotidian objects as an essential part of the process of aesthetic impression formation in daily life. In our system, each piece of furniture has a functional value, that is, it serves a purpose. For example, the purpose of a chair is that a person sits on it. A composition might affect the functional value of one or more pieces of furniture. For example, if a chair is situated on the bed, neither the bed nor the chair can serve its purpose anymore. When the functional value of a piece of furniture is altered, a Tension of Functional Value is triggered.

     
  5. 5.

    Distribution. Lidwel et al. (2005) affirm that symmetry has always been associated with beauty. Although we believe that asymmetrical distributions can also be interesting, in order to implement this computer program we have decided that in this work furniture must be uniformly distributed around the room producing a symmetrical arrangement of objects. Otherwise, a tension of distribution is triggered.

    Lidwel et al. (2005, p. 172) suggest that we human beings have a preference for spaces resembling a savannah landscape than for spaces resembling simple landscapes like a desert, or dense and complicated landscapes like jungles or mountains. They suggest that the characteristics that we like the most about the savannah are the open spaces, scattered trees, and a uniform meadow; these in opposition to obstructed views, unordered complexity and irregular textures (e.g. see Balling and Falkin 1982; Kellert 1993). Inspired by these comments we establish our last two rules.

     
  6. 6.

    Density. In this work, an interior design composition must include a balanced amount of elements. Thus, if the number of household goods inside the room is too small, resembling a simple landscape, a tension due to low density is triggered. On the other hand, if the number of household goods inside the room is too big, resembling a dense or complicated landscape, a tension due to high density is triggered.

     
  7. 7.

    Proximity. In our model, the distance between each individual piece of furniture must be at least 10 cm (with the exception of the seats located by the dining-room table). Thus, when two or more pieces of furniture are too close to each other (less than the equivalent to 10 cm) the tension of proximity is triggered.

     

We do not suggest that these rules are complete (e.g. we have not taken into consideration the effect of light on colours) or that these are the only rules that designers might use. We have defined them in order to develop our computer program. The outcomes of our program will shed some light about their utility and how they can be complemented.

Implementation of the ERI-Designer

The ERI-Designer model was built on top of a software called Sweet Home 3D version 1.5.1 developed by Emmanuel Puybaret (2009). It includes a 2D working area and a 3D viewer area (see Fig. 1 a, b). In order to implement the ERI-Designer, we decided to employ a virtual room of 6 × 6 m as a setting for the interior design composition. The room includes one window and one door. The size of the room, placement of the window, placement of the door, and the colour and pattern of the floor are fixed, i.e. they cannot be modified.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig1_HTML.jpg
Fig. 1

Working area for the interior design composition

There are 60 types of furniture, divided in 7 furniture-groups, which can be employed to create the composition (see Fig. 1c): group 1 includes 9 types of beds; group 2 includes 7 types of household goods to store objects; group 3 includes 13 types of seats; group 4 includes 7 types of sofas; group 5 includes 6 types of coffee tables; group 6 includes 5 types of desks; group 7 includes 13 types of dinner room tables. The colours (or hues) of walls and furniture are modifiable. ERI-Designer employs seven hues: blue, green, purple, yellow, orange, red and pink. “Black, white, and grey lack hue and are considered neutrals[...] [Colour]value is the lightness or darkness of a surface color.” (Fehrman and Fehrman 2004, p. 8). In this work each hue has three values: light, regular and dark. Thus, we have 7 hues with three colour-values for each (that gives us 21 possibilities). The system also employs five neutrals: black, white and three types of grey. Thus, the ERI-Designer includes 26 options to assign a colour to the different elements that comprise a composition.

The Observation Module

The ERI-Designer includes a module called observation module. The purpose of this module is to allow the system to recognize situations that trigger affective reactions and to record information that might be useful for the development of novel designs. The observation module is capable of: determining the hue, colour-value, position and orientation of every piece of furniture inside the room; grouping household goods based on their proximity; grouping household goods based on their colour; determining the use or function of the room (living-room, dining-room, bedroom and study room); assigning to proximity groups one of four possible shape-labels: tendency to form a circle, tendency to form a triangle, tendency to form a square, tendency to form a line, and so on.

The routines inside the observation module are known as the observation process. The observation process works as follows (see Fig. 2). The system registers all data it can obtain from the room in 3D. However, not all these information is useful for the composition. The observation process focuses only in those elements that are relevant for the task in progress. So, the system filters all irrelevant information creating a new structure known as the Context. Then, the system runs the abstraction process. This process employs the Context and the CID-rules to create a new structure known as the Tensional-Context that represents the room in terms of tensions and groups. If the user modifies the room the observation process can be run again and the context and Tensional-Context are updated with the new information. During the creation of knowledge structures the user activates the observation process; during the creation of novel designs through E-R cycles the system itself activates the observation process.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig2_HTML.gif
Fig. 2

The observation module

Representation of Tensions in the ERI-Designer

We have defined seven variables representing the seven tensions described by the CID-Rules. Each one of them has a name associated and a mnemonic (see Table 1): (1) Tension due to colour harmony (Tch); (2) Tension due to colour contrast (Tcc); (3) Tension due to colour impression (Tci); (4) Tension due to functional value (Tfv); (5) Tension due to distribution (Tdi); (6) Tension due to density (Tde); and (7) Tension due to proximity (Tpr). All tensions have three possible values: 0, +1, +2. Additionally, the tension due to colour contrast (Tcc) and the tension due to density (Tde) might have two negative values: −1 and −2.
Table 1

Tensions values and mnemonics

Name of the tension

Mnem.

Possible values

Colour harmony

Tch

0, +1, +2

Colour contrast (monotony and excessive contrast)

Tcc

−2,  −1,  0,  +1,  +2

Colour impression

Tci

0, +1, +2

Functional value

Tfv

0, +1, +2

Distribution

Tdi

0, +1, +2

Density (low and high density)

Tde

−2,  −1, 0,  +1,  +2

Proximity

Tpr

0, +1, +2

The absolute value of the variable indicates the intensity of the tension: 0 represents the fact that the observation module does not detect the conditions required to trigger a tension; 1 represents the fact that the observation module detects the conditions needed to trigger a tension; 2 represents the fact that the observation module detects the conditions needed to trigger a high tension. In the case of the tension due to colour contrast, a negative value indicates that it is triggered due to monotony, while a positive value indicates that the tension is triggered due to an excessive contrast. In the case of the tension due to density, a negative value indicates that the tension is triggered due to low density, while a positive value indicates that the tension is triggered due to a high density.

In the following lines we explain the procedures that the observation module employs to analyze the composition.
  1. 1.

    Tension due to colour harmony. In order to calculate the tension due to colour harmony, the system groups together all pixels sharing the same colour. If the system finds that all colours are adjacent, then the variable tension due to colour harmony is set to zero; if the system finds that the colour of one or more groups are separated by one position from the colour of the reference group, the variable tension due to colour harmony is set to +1; finally, if the system finds that the colour of one or more groups are separated by two or more positions from the colour of the reference group, the variable tension due to colour harmony is set to +2. In this work, when white, black and grey are combined with any other colour the system always rises a tension due to colour harmony equal to +1.

     
  2. 2.

    Tension due to colour contrasts. In this work, two colours contrast when: (1) The value of one of them is light and the value of the other one is either regular or dark; (2) the value of each colour is either regular or dark but their hues are different.

     
  3. 3.

    Tension due to colour impression. The system first determines the predominant colour within the composition employing the same procedure described earlier. Then, it employs the following rules to set the variable tension due to colour impression: a) if the value of the predominant colour is light then the tension is set to zero; b) if the value of the predominant colour is regular then the tension is set to +1; c) if the value of the predominant colour is dark then the tension is set to +2.

     
  4. 4.

    Tension due to functional value. The observation module analyses each piece of furniture in the room and verifies if its functionality has been affected.

     
  5. 5.

    Tension due to distribution. The room is divided in 9 locations: top left corner (TL), top centre (TC), top right corner (TR), centre left (CL), centre (CC), centre right (CR), bottom left corner (BL), bottom centre (BC) and bottom right corner (BR) (see Fig. 3a). Based on these locations, the system verifies if objects are distributed symmetrically around the room. If ERI-Designer finds one asymmetrical distribution the tension is set to +1; if the system finds two or more asymmetrical distributions the tension is set to +2. For example, in Fig. 3b the system triggers a tension due to distribution equal to +1 while the tension in Fig. 3c is equal to zero because the room has been balanced.

     
  6. 6.

    Tension due to density. The system calculates the percentage of area covered by furnishing inside the room. Then, it employs Table 2 to assign the right value to the tension due to density. All percentages in Table 2 can be modified by the user.

     
  7. 7.

    Tension due to proximity. The system calculates how many pieces of furniture are located too close to other objects. It employs Table 3 to determine the value of the tension due to proximity. All percentages described in Table 3 can be modified by the user.

     
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig3_HTML.jpg
Fig. 3

Nine locations inside the room and examples of furniture located in the room

Table 2

Values of the tension due to density

Percentage of area covered by furniture (%)

Tension due to density

0–10

−2

10–20

−1

20–50

0

50–65

1

65–100

2

Table 3

Values of the tension due to proximity

Percentage of furniture located too close to other objects (%)

Tension due to proximity

0–25

0

26–74

+1

75–100

+2

Creation of Knowledge Structures

In order to create the knowledge structures ERI-Designer provides an interface that represents the room described earlier and the 60 types of furniture. The user employs the interface to generate arrangements of furniture. The system records each step in the process and uses this information to create its knowledge-base. In this way, the system has the capacity to build its knowledge structures from the experience and knowledge of humans. The examples provided by the user to create the system’s knowledge-base are known as the Previous-Designs.

In this work we define three classes of actions: the basic-action, the macro-action and the generalized-action. The following lines describe the first two types of actions; the third type is described some lines further down. Each piece of furniture has several attributes associated with it: a number that works as a unique identifier, its type, position, orientation, hue and colour-value. Basic-actions update these attributes. The user can perform any of the following basic-actions: put a piece of furniture on any of the defined positions; eliminate a piece of furniture from the room; move a piece of furniture from one position to a new position; change the orientation of a piece of furniture; change the hue and colour-value of each piece of furniture. Macro-actions are comprised by one or more basic-actions. Macro-actions allow manipulating furniture-groups. For simplicity, from now on, macro-actions will be referred to as actions.

The ERI-Designer requires that the user indicates to the system which sequences of basic-actions form one action. Then, the system updates the context and Tensional-context structures. These three processes, the creation of an action and the updating of both structures, are executed when the user activates or triggers the observation process. The following lines provide an example. The user starts the composition locating some furniture in the room. For instance, let us imagine that the user locates a dinner table and four chairs on the position top-right. That is, the user performs five basic-actions: move dinner-table 1, move chair 1, move chair 2, move chair 3 and move chair 4 to the selected position. At this point the user activates the observation module. As a consequence three things happen: (1) the system comprises the five primitive-actions into one single macro-action known as action-1; (2) the context is updated; the content of the context at this point is referred to as context-1; (3) the Tensional-context is updated; the content of the Tensional-context at this point is known as Tensional-context-1. That is:
$$ \begin{aligned} {\tt Action\,1\,\hbox{->}\,partial{\hbox{-}}}&{\tt composition{\hbox{-}}1}\\ &{\tt [context\,1]\,Tensional{\hbox{-}}context\,1} \end{aligned} $$
In this project, a composition is the result of a sequence of actions or operations performed by the designer. Each time an action is executed, either a new element is included within the composition or existing elements are modified or eliminated. Thus, we can describe a composition as a process that consists on sequentially applying a set of actions, which generate several partial or incomplete works and their corresponding contexts and Tensional-contexts, until the right composition arises or the process is abandoned. Now we are in a position to explain how the system creates its knowledge structures. The procedure has five steps and works as follows:
  1. 1.

    The user performs some basic-actions and then triggers the observation process; so, a new action is created and the context and the Tensional-context are updated. That is, action-1, context-1 and Tensional-context-1 are generated. In this case, action-1 consists on locating two orange sofas and one orange coffee-table at position TL and painting the walls in yellow (see Fig. 4a). Now, the Context is updated generating Context-1, which includes the following information:

    Context-1, general information about the room:
    • The predominant colour-value and hue inside the room are light-yellow (all furniture and walls are taken into consideration to calculate this value).

    • The room is classified as a living-room

    • The distribution of furniture is asymmetrical

    • The furniture covers less than 10% of the room’s area.

    Context-1, information about groups:
    • The system detects one proximity group. The information about the group includes its shape: tendency to form a triangle; its position: TL; and the unique identifier of each piece of furniture that forms the group.

    • The system also detects one colour group and registers it in the context including the same information just described

    • Context-1, information about walls and furniture:

    • The colour-value and hue of walls are light-yellow.

    • The predominant colour-value and hue of all furniture is standard-orange.

    • The system also calculates the second predominant colour-value and hue of all furniture. In this case, such values do not exist.

    Context-1, detailed information about furniture:
    • The context includes information about each piece of furniture in the room. For example, the hue and colour-value of sofa-1 is light-orange; coffee-table-1 is located in front and close to sofa-1 (between 21 and 50 cm); wall-4 is located next to the back of the sofa-1 (less than 10 cm); wall-1 is located next to the left side of the sofa-1 (less than 10 cm); and so on. For the sake of clarity, the rest of the detailed information about furniture is excluded from this example.

    With this information the system updates the Tensional-context. The Tensional -context is formed by the seven tensions described earlier, plus a slot that specifies the function of the room and a second slot that indicates the number of proximity groups inside the room. Thus, for this example, the Tensional-context has the following values: Tension due to colour harmony: 0; Tension due to colour contrasts: 0; Tension due to colour impression: 0; Tension due to Functional Value: 0; Tension due to distribution: 1; Tension due to density: −2; Tension due to proximity: 0; Function of the room: living-room; Number of proximity groups: 1. Figure 5 shows a graphical representation of the Tensional-context-1. The image of two sofas indicates that the room has been classified as living-room. The dot inside the dashed-square indicates that the room includes one proximity group.

     
  2. 2.

    The user performs more basic-actions. Then, it triggers the Observation process and action-2, context-2 and the Tensional-Context 2 are generated. In this example, action-2 is formed by five basic-actions: the user inserts one dinner-room table and four chairs at the position BR, whose colour-value and hue are all equal to dark-green (see Fig. 4b).

     
  3. 3.

    The system generalizes the last action performed. Actions include detailed information about each piece of furniture they modify: its unique identifier, type, position, orientation, hue and colour-value. This precise information is important to progress the current composition but probably useless if one wants to employ this information in other designs. So, the system generalizes the action in order to register the essential information about it.

     
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig4_HTML.jpg
Fig. 4

Different stages of the composition during the creation of knowledge structures

https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig5_HTML.gif
Fig. 5

Tensional-Context-1 (Atom-1) produced by the first action

The generalization consists of determining if the action inserts, eliminates or modifies furniture in the composition. In this paper we only use actions that insert furniture, i.e., actions that insert proximity groups comprised by furniture. Then, the system calculates how many proximity groups are inserted into the composition by the action, how many members each proximity group has and the furniture-group that each member of each proximity group belongs to. For our current example, action 2 inserts one proximity group comprised by one member of furniture-group 7 (dinner-room tables) and four members of furniture-group 3 (seats). That is, the action inserts a dining-room suite. The system determines the value of the shape-label of the group; in this case, the shape-label indicates a tendency to form a circle.
As part of the same process, the system analyses the effect that the action has in the distribution and colour impression of furniture inside the room. The idea is to represent what we refer to as the distribution and colour “intention” of the action. Thus, the system analyses if action-2 increases, decreases or does not modify the tension due to distribution in the room and register this information. In the same way, the system analyses if action-2 involves incorporating new furniture in the composition. If it does, the system checks if the hue and colour-value of the new furniture match the hue and colour-value of the predominant or second predominant furniture in the room. The system attempts to determine if the “intention” of the action is to keep on using the same furniture colours or incorporating novel colours into the composition. The system registers this information. Thus, in summary, the generalized action-2 can be described as follows:
  • The action inserts into the composition one proximity group comprised by one member of furniture-group 7 (dinner-room tables) and four members of furniture-group 3 (seats). That is, the action inserts a dining-room suite.

  • The shape-label of this group is equal to tendency to form a circle.

  • This action must attempt to decrease the tension due to distribution.

  • The hue of the members of the group must be different to the hue of the predominant furniture colour in the room (in this example, the predominant furniture hue is orange).

  • The colour-value of the members of this group must be equal to the colour-value of the predominant furniture colour in the room (in this case the colour-value is dark).

  1. 4.

    The system copies the content of the Tensional-context-1 into a new structure in memory known as atom-1. This is necessary because the system constantly updates the Tensional-context and if we do not save the information it is lost. Then, ERI-Designer associates atom-1 to the generalized action-2. Thus, the system records that, when a composition in progress can be represented in terms of tensions and groups equal or similar to those in the structure Tensional-context-1, something logical to do in order to progress the composition is to perform the generalized-action-2. generalized-actions represent more a kind of sketch, a rough outline of the possibilities to continue the design process than concrete actions. When ERI-Designer generates novel compositions, during reflection it employs a set of routines to concrete these generalized actions.

     
  2. 5.

    Now the user performs more basic-actions. Then, it triggers the observation process and action-3, context-3 and the Tensional-context-3 are generated. In this example, action-3 inserts two green sofas at location TR (see Fig. 4c). Next, the generalized-action-3 is produced. Then, employing the Tensional-context-2 the system generates atom-2 and associates to it the generalized-action-3. The process continues until the user ends the design process. Thus, the last atom created has associated the special generalized-action known as end-of-composition.

     

Development of a Composition

ERI-Designer produces novel compositions as a result of cycles of engagement and reflection: during engagement the system chooses one generalized-action to progress the composition; then, it switches to reflection to evaluate and, if it is necessary, modify the composition generated so far. Then, it switches back to engagement and the cycle continues. Engagement works as follows:
  1. 1.

    The user provides an initial state. This initial state, together with the system’s constraints (fixed size of the room, fixed position of the window and the door, and fixed type and colour of the floor), are considered the initial requirements of the design.

     
  2. 2.

    The ERI-Designer updates the context and the Tensional-context of the current composition.

     
  3. 3.

    The system employs the Tensional-context structure as cue to probe memory and matches all atoms that are equal or similar to such a structure. Then, it retrieves all generalized-actions associated to each matched atom. In this work, a Tensional-context is considered as similar to an atom when the former is at least 50% equal to the latter. This percentage is known as the ACAS-constant and can be modified by the user. As it will be shown later, this constant is important for the model.

     
  4. 4.

    The system selects at random one action between the available retrieved generalized-actions. Then, the action is executed modifying the composition (as mentioned earlier, for the sake of clarity, all generalized-actions in this paper insert proximity groups in the composition).

     
At this point the system switches to reflection, where it performs three routines:
  1. 1.

    Compression of the proximity group. Because generalized-actions do not include precise information about the position of each piece of furniture, the system might locate them inside the room in a kind of loosed way. The system attempts to solve this problem by reducing the area occupied by the proximity group. That is, it performs small modifications to the orientation and position of each piece of furniture inside the proximity group until the group is organized in a more compact and ordered way (see example behind).

     
  2. 2.

    Homogenize the type of furniture. During engagement, the system selects at random the types of the furniture that it inserts in the composition. Thus, the system might include in the same proximity group inharmonious types of furniture. So, during reflection the system attempts to solve this problem by selecting at random one element inside the proximity group as a reference, and assigning to the rest of them the same type as the reference. The system includes a parameter defined by the user that determines the probability that this process is performed only partially, that is, that some pieces of furniture keep its original type. The purpose of this parameter is to allow the system to generate unexpected combinations of types of furniture.

     
  3. 3.

    Uniformity of colour. During engagement, the system selects at random the colours of the furniture that it inserts in the composition. Remember that the generalized-action only includes information like “The hue of the members of the group must be different to the hue of the predominant furniture colour in the room”. Thus, the system might assign completely different colours to each member of the proximity group. The system attempts to solve this problem by selecting at random one element inside the group as a reference and assigning to the rest of them the same hue and colour-value as the reference.

     

The system includes a process that attempts to establish a colour-link with other proximity groups inside the room. It works as follows. The system selects at random one proximity group that already was part of the composition. Then, it selects also at random one of its elements and uses its hue and colour-value as reference. Then, the system assigns to two of the elements inside the new proximity group the same hue and colour-value as the reference. The system includes a parameter defined by the user that determines the probability that this process is performed.

Then, the system switches back to step 2 in engagement and the process continues. The cycle ends when the system matches an atom that has associated the special action end-of-composition. The following lines present an example of a novel composition.

A Thorough Example of the Composition Process

In this example the ACAS-Constant is set to 50%. We employ a knowledge base that comprises 120 atoms. The user provides the initial state showed in Fig. 6.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig6_HTML.jpg
Fig. 6

Three different views of the initial state

This partial composition generates the Tensional-context-1 showed in Fig. 7. The tension due to colour harmony has a value equal to one because the walls are white (remember that by default, white, gray and black trigger a tension of colour harmony equal to one). The tension due to colour contrast is set to −1 because the colour-value of the sofa and the wall are equal to light. The tension due to distribution is set to one because there is an asymmetrical distribution of furniture. The tension due to density is equal to −2 because the room is almost empty. Engagement starts and the Tensional-context-1 matches 12 atoms. It selects one at random and executes its associated generalized-action:
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig7_HTML.gif
Fig. 7

Tensional-Context-1

Inserts into the composition one proximity group comprised by one member of furniture-group 7 (dinner-room tables) and four members of furniture-group 3 (seats). That is, the action inserts a dining-room suite. The types of the seats and the type of the dinner-room table are selected at random. The shape-label of this group is equal to tendency to form a circle. So, the system attempts distribute the seats around the table. This action must attempt to decrease the tension due to distribution. The system attempts to look for a position that decrements the tension. If the system cannot find such a position it locates the group in any available location. In this case, the system locates the new group in the position BL and decreases the tension. The hue of the members of the group must be equal to the hue of the predominant furniture colour in the room. In this example, the predominant furniture hue is yellow. The colour-value of the members of this group must contrast the colour-value of the predominant furniture colour in the room. In this example, the predominant colour-value is light, so the colour-value of the members of the group is set to standard or dark.

This produces the composition showed in Fig. 8. Notice that the seats have different types and they are not well arranged around the table.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig8_HTML.jpg
Fig. 8

Different views of the partial composition-1

Now, the system switches to reflection. The system initiates the routine Compression-of-the-proximity-group. So, it starts to rotate the elements that comprise the proximity group (to modify its orientation and position) in order to organize the group in a more compact and ordered way. Figure 9a, b and c show the changes in the organization of the group performed by the routine.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig9_HTML.jpg
Fig. 9

The Compression-of-the-proximity-group routine organizes the group in a more compact and ordered way. Images (a), (b) and (c) show the modifications

Now, the second routine, Homogenize-the-type-of-furniture, is run. Its function is to homogenize the types of furniture in the group. So, the system selects seat one as the reference, and the type of the rest of the seats is equalized to the type of the reference (see Fig. 10a). The next step is to run the routine Uniformity-of-colour. So, the system selects as a reference the hue and colour-value of the table. So, the hue and colour-value of all seats are equalized to the colour-value of the reference (see Fig. 10b). Finally, the system triggers the process to generate colour-links between groups (in this example, the probability of triggering this process is equal to 30%; this percentage can be modified by the user). In this case, the hue and colour-value of two seats are equalised to the colour and value of the sofa (see Fig. 10c).
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig10_HTML.jpg
Fig. 10

Three routines performed during reflection. Image a shows the homogenization of the types of seats; b shows the homogenization of colour; c shows the creation of a colour-link

Now the system switches to engagement and the context and Tensional-context is updated. Figure 11 shows the Tensional-context of the current composition.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig11_HTML.gif
Fig. 11

Tensional-Context-2

The room is now classified as living-room/dining-room (the image with the dining-room table indicates that the room has been classified as a dining-room). It includes two proximity groups (that is why the dashed-box includes two points). The colour-value of most furniture stills light; so, the tension due to contrast keeps its value of −1. With the inclusion of the new group at the position BL the tension due to distribution is set to zero (instead of 1). Because there are more pieces of furniture inside the room, i.e. the room is not that empty, the tension due to density decreases its value to −1. The tension due to proximity is set to +1 because the seats are too close to each other in the room. Now, the system looks for an atom that is equal or similar to the Tensional-context-2. It matches seven atoms, selects one at random and performs its generalized-action. It consist of inserting 4 sofas and one coffee-table, and the action must attempt to decrease the tension due to distribution. The system attempts to find a position in the room that decreases the tension due to distribution. However, any empty position will increases it. So, the system selects a location at random (see Fig. 12a). The system switches to reflection. The routine Compression-of-the-proximity-group starts and rearranges the group (see Fig. 12b, c shows a different view of the same image). The second routine, Homogenize-the-type-of-furniture, does not change the composition because the sofas belong to the same group. Something similar happens with the third routine Uniformity-of-colour: because all members of the group share the same hue and colour-value the composition is not modified. Finally, there is a 30% probability of triggering the process to create colour-links; in this occasion the system does not start it.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig12_HTML.jpg
Fig. 12

Different views of the partial composition-2

Now, the system switches back to engagement and the context and Tensional-context are updated. Figure 13 shows the Tensional-context-3.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig13_HTML.gif
Fig. 13

Tensional-context-3

Now, the system looks for an atom that is equal or similar to the Tensional-context-3. It matches nine atoms, selects one at random and performs its generalized-action. It basically consists of inserting a desk and two seats. Figure 14a shows the room after the action has been performed.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig14_HTML.jpg
Fig. 14

Partial composition-3

Then, the system switches to reflection. The routine Compression-of-the-proximity-group rearranges the inserted group. Figure 14 a, b and c show how the group was ordered.

The second routine, Homogenize-the-type-of-furniture, selects the seat lacking the back as a type-reference and changes the chair’s type into the same type as the reference (see Fig. 15a). The third routine, Uniformity-of-colour, selects the dark-green as a reference and assigns these colour-value and hue to the whole group (see Fig. 15b). Finally, the process to create colour-links is triggered; so, the desk gets the same hue and colour-value that the table in the proximity group 2 (see Fig. 15c). Now, the system switches back to engagement and the context and Tensional-context are updated. Figure 16 shows the Tensional-context-4. The function of the room is classified as: living-room/dining-room/study room (the image with the desk indicates that the room has been classified as a study room).
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig15_HTML.jpg
Fig. 15

Modifications performed during reflection to the partial composition-3

https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig16_HTML.gif
Fig. 16

Tensional-context-4

Now, the system looks for an atom that is equal or similar to the Tensional-context-4. It matches nine atoms and selects one at random. This time, the generalized-action indicates to finish the design and the E-R cycle stops. Different views of the final design are shown in Fig. 17.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig17_HTML.jpg
Fig. 17

Different views of the final composition

Evaluation

ERI-Designer has been evaluated by means of an Internet questionnaire. Fifty subjects from six different countries took part: 74% were Mexicans, 10% were British, 6% were Germans, 4% were Spanish, and 2% were French and Colombians (2% did not provide information about their origin). The average age of the participants was 36.39 years of age; the youngest was 21, the oldest 68. 72% were females and 28% were males. The subjects were informed that the purpose of the study was to evaluate some aspects of interior design products generated by human and computational agents. The questionnaire showed three different views of each of three computer-generated representations of rooms. Room A was generated by ERI-Designer with the ACAS-Constant set to 50% (see Fig. 17); Room B was copied from an interior design book (it was designed by a human professional), adapting it to the restrictions of the experiment described some lines ahead; Room C was generated by ERI-Designer with the ACAS-Constant set to 30% (see Fig. 18). The subjects did not know which of them were designed by a computer agent and which of them were designed by a human. The questionnaire explained that some characteristics of the rooms were fixed by the research team and could not be modified: its size (6 × 6 m), placement of the window, placement of the door, the colour and pattern of the floor. Other characteristics were determined by the designers working on the task: the colours of walls and furniture as well as the type, position and orientation of each piece of furniture inside the room. The use of relatively small decorative elements like pictures, plants, vases, and so on, was not allowed at this stage. For each room, subjects were asked to answer the following five questions on a five point scale (from “not at all” to “very much”): Do you like the colours employed in the room (only walls and furniture)?; Do you like the distribution of the furniture inside the room?; Do you consider that the elements that comprise the room (only walls and furniture) combine harmoniously with each other?; Would you feel comfortable if you had to use this room?; Would you classify this room as original?
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig18_HTML.jpg
Fig. 18

A design copied from a book (top); a design produced by the system with the ACAS-Constant set to 30% (bottom)

Subjects were also asked to choose, between the following 5 possible answers, who had designed the room: 1. Recreational interior designer; 2. Student of interior design; 3. Senior/student of interior design; 4. Professional interior designer; 5. Senior/professional interior designer. The questionnaire included a space to write free comments about each room. In the last part of the questionnaire, subjects were asked to rank the three rooms according to their personal preference by simply ordering them from best to worst. It was hypothesized that Room B (copied from an interior design book) would gain the highest rates, that Room A (ERI-Designer with ACAS-Constant set to 50%) would gain slightly lower rates than Room B, and that Room C (ERI-Designer with ACAS-Constant set to 30%) would gain the lowest rates.

Results

Following Greene and d’Oliveira (1981) values from 1 to 5 are allotted to the five-point scale between “not at all” and “very much”. For each evaluated aspect, the assessment of each room is equal to its mean. Figure 19 shows the results of the first five questions. The vertical axis plots the mean of the answers provided by the subjects and the horizontal axis the five characteristics to be evaluated. As expected, Room C obtained the lowest rates, and Room A and B obtained very similar rates in all aspects but originality, where clearly Room B was perceived as more original than the other two.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig19_HTML.gif
Fig. 19

Evaluation of colour, distribution, harmony, comfort and originality in each of the three rooms

Surprisingly, Room A was slightly higher evaluated than Room B. From these results it was not possible to conclude that ERI-Designer was a better designer than a human author, but perhaps that we were less effective in accomplishing the task of representing a successful and somewhat contrived room in computer terms. Nevertheless, it was encouraging that Room A obtained similar rates to those obtained by Room B. This situation suggested that we were in the right track. Room C got the lowest rates in all the evaluated aspects. The comparison between Room A and Room C’s results allowed observing the importance of the ACAS-Constant for the ERI-Designer: low values of the ACAS-Constant produced arrangements that were perceived as inharmonious; i.e., it seemed that subjects disliked the colours employed in the composition, as well as the selection and distribution of the furniture (this claim was supported by the free comments made by some of the participants). As a consequence, Room C was classified as uncomfortable. In this way, low values of the ACAS-Constant generated poor compositions. When the ACAS-Constant was set to 50% the system produced its best compositions. When the ACAS-Constant was set to 100% the system only reproduced the sequences of actions recorded in its data-base. These results coincided with those reported by Pérez y Pérez (2007) in MEXICA, an automatic plot-generator based on the E-R model.

The next question in the questionnaire asked subjects to choose, between 5 possible answers, who had designed each room. Figure 20 shows the results. The vertical axis plots the percentage of subjects that answered the question, and the horizontal axis plots the five possible answers. 26% of the subjects resolved that Room A was designed by recreational interior designer, 44% determined that it was designed by a student of interior design, and 20% opted by a senior/student of interior design. 42% of the subjects resolved that Room B was designed by a recreational designer, 32% determined that it was designed by a student of interior design, and 12% opted by a professional designer. A 68% of the subjects resolved that Room C was designed by a recreational designer and 18% opted by student of interior design. Due to the constrictions of the study, it was not surprising that most subjects concluded that the three rooms were designed either by recreational designers or by students of design. There was a tendency to identify Room A with the work of a student. Surprisingly, most subjects identified Room B with the work of a recreational designer, although a significant amount of participants also identified it with the work of a student. Nevertheless, only Room B was considered the work of a professional designer by more than 10% of the partakers. Finally, Room C was considered by a forceful amount of subjects as the work of a recreational designer.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig20_HTML.gif
Fig. 20

Types of designer selected by the subjects

Finally, participants ranked the three rooms according to their personal preference by simply ordering them from best to worst. Figure 21 shows these results. The vertical axis plots the percentage of subjects that answered the question, and the horizontal axis shows their ranking order. There was a correlation between this ranking and the previous results. Although Room A got the highest rank, Fig. 21 does not show a clear cut preference for either Room A or Room B. By contrast, subjects unambiguously considered Room C as the worst of them all.
https://static-content.springer.com/image/art%3A10.1007%2Fs11023-010-9208-9/MediaObjects/11023_2010_9208_Fig21_HTML.gif
Fig. 21

Ranking of the three rooms

Discussion

The purpose of the first version of the ERI-Designer is to evaluate if the E-R model might be useful to build a computational version of the interior design process. This requires creating an adequate representation of the knowledge necessary to develop an interior design composition that can be employed by the engagement-reflection model. After studying how designers design, and inspired by their experience’s accounts, we propose a set of rules that determine the characteristics that our computational agent perceives from its environment. These characteristics are employed to build a knowledge structure know as the Tensional-Context, which drives the generation process during the ER cycles. In this way, the Tensional-Context is a computational representation of the environment in terms of the tensions and the groups of elements that comprise the room. As far as we know, there is no similar computer representation of knowledge for design. In this way, ERI-Designer allows studying some of the effects that the environment has in the creative process of a computational agent. The first outcomes generated by our system suggest that the Tensional-Context is an adequate knowledge structure to drive the composition process. Our plans for future versions of the system contemplate research on how to generate more complete knowledge representations. For example, we are interested in studying how to represent explicit relations between the groups of furniture inside the room, and how these affect the generation process.

The ER Model includes two main processes: the construction of its knowledge-base from a set of well designed examples provided by the user; the generation of new compositions. The graphical interface of ERI-Designer provides a good intuitive mechanism for the non-technical user to produce the set of examples that the system requires. A log file is created in order to have the possibility of analyzing all the knowledge-structures built by the system.

ERI-Designer is capable of generating outputs that some people think adequate. For example, some compositions developed by the system are considered relatively harmonious by a group of human judges. This characteristic arises as a result of the E-R cycles. That is, instead of using optimization processes like other similar systems, ERI-Designer employs the Tensional-Contexts to retrieve actions that produce correct combination of elements within a composition. This is important for two reasons. First, we are interested in contributing to the understanding of human creativity and we believe that a cognitively-inspired model is more useful to achieve this goal than an engineering system. Secondly, because a core goal of this work is to study if the ER-Model can be used in different domains, the production of adequate outputs employing Tensional-Contexts are essential for this research. A significant part of the information that determines what is a correct or harmonious combination of elements is encoded in the atoms (although some information is also defined in some routines performed during Reflection). Because atoms are built from the set of Previous-Designs, ERI-Designer can modify, relatively easily, the steps that it follows to generate a correct combination of elements; i.e., it can easily work with different design styles. Some of the requirements and constraints of the design task are explicitly defined before the ER cycle starts (e.g. the size of the room, position and size of the window and the door, and so on); however, some others emerge during the composition process (harmonious use of colours, type of furniture, distribution, etc.). The routines that drive the selection and arrangement of colours and furniture in the system can adapt to such dynamic constraints that arise as a result of the composition process itself. Again, all this flexibility is an important characteristic of the system.

The current version of the system selects at random the next action to perform. This characteristic provides flexibility to the system and allows emerging unexpected designs. All actions associated to the same atom share the same Tensional-Context. Thus, although their content might not been the same, and therefore they might lead to different directions, all they represent logical possible actions to perform under the current state of affair of the composition. Nevertheless, future versions of the system will incorporate routines that select the next action to perform based on the current necessities of the composition.

The evaluation of the system suggests that the ER-Model might be employed to represent the interior design process. Although the design produced by a human is taken out from its original context, and therefore it cannot be considered as a complete design product, it is encouraging that the room generated by the first version of ERI-Designer got the best rates and it was chosen as the best room. However, as expected, the evaluation also shows that much more work is required. The highest rate obtained by ERI-Designer is 3.12 points, and the average is 2.59 points. These results clearly suggest that subjects do not consider the output produced by the system as a good design. Furthermore, its output was considered by most participants as the product of a student of interior design (just one point above than the lowest possible evaluation). Although the designed employed for the questionnaire was novel in the sense that it was not present in the system’s knowledge-base, it was perceived by the subjects in the study as lacking originality. Therefore, it is necessary to improve the routines to generate novel outputs.

ERI-Designer incorporates characteristics not present in previous models. For example, sub-symbolic systems such as neural networks or genetic algorithms have proven to be effective in classification problems and optimization. In some of them, random variables are used to introduce unexpected behaviour in the system. Although these can be adjusted to produce some interesting outcome, it is usually difficult to trace the exact reasons why the outcome was produced, that is: build an explanation for the successful solution. Our system reasons with representations of affective reactions associated to actions to be performed, which constitute a good level of abstraction to build explanations. Sub-symbolic systems are successful in optimization problems where the process and the goals are well known in advance but arriving at a solution is just too much work. The coding and evaluation function are more effectively established when the objective of the problem is well specified and it is clear what a good solution is (cf. Michalewicz 1999). The design process is not a well established, well known, standard methodology, like a science. It is rather an empirical trial-and-error process, based on experience, which usually allows for several solutions to be acceptable to a given problem (Maher et al. 1988). Therefore, the successful use of sub-symbolic techniques depends very much on finding suitable coding schemes and evaluation functions to carry the process through in a sound way. Results can certainly be very impressive, but they depend on a clear definition of problem and solution(s). By contrast, ERI-Designer is flexible. It uses the experiences encoded in the Previous-Designs to build its knowledge-base, which can easily been modified. The present version of the system does at engagement a selection of next actions in terms of furniture units (dining room, lounge, studio, etc.). That is, during the idea-association the system builds a general sketch, a global solution without paying attention to the details: the kind of table, colour, etc.; it is during reflection that the adequacy of the solution is assessed and the details are developed. This mechanism adds an extra feature to the model that mimics the way humans do top-down reasoning during the design process. The tension-based routines used in the system also give the versatility to scale up or down the granularity at which the elements of the design process (furniture) can be represented and reasoned about. It could be possible, for instance, to customize the model to design buildings where different type of rooms are generated; or, conversely, narrow the scope of detail down to the level of parts of furniture and decide to use the system to design chairs, tables, etc. All these different levels would follow the exact same logic and the system would provide a rationale for the decisions taken in terms of tensions. This would allow building a system that can work with different elements at different levels of granularity within the same composition. These ideas could lead to possible extensions to the system that we have yet to investigate.

Summing up, the system we present in this paper is based on engagement-reflection, a model used to implement systems to simulate computational creativity. As far as we know, our system is the first one to apply such a model to design. This first prototype shows the plausibility of employing the E-R model to represent part of the process of interior design. But we are aware that this is only the first step and much more work is needed. We hope this work encourage other researchers to participate in this effort.

Copyright information

© Springer Science+Business Media B.V. 2010