GENERATIVE DESIGN EXPLORATION FRAMEWORK
BASED ON SUBJECTIVE EVALUATION
Jia Cui and Ming Xi Tang
School of Design, The Hong Kong Polytechnic University, Hong Kong, Hong Kong
Keywords: Generative system, Design exploration, Evolutionary algorithm, Subjective evaluation.
Abstract: Application of evolutionary computation in design, supporting artificial intelligence and design inspiration,
requires a good understanding of design process. However, obtaining a creative design solution with
subjective evaluation is the barrier of traditional evolutionary mechanisms. In this paper, a novel aesthetic
evaluation model connecting subjective and objective space is introduced, and an exploration algorithm
combining human cognition and preference is presented, which can support design exploration to generate
new design solutions more effectivelly and intelligently.
1 INTRODUCTION
In AI in Design community, it is considered that
complex human intelligent activities being reduced
to manageable computing task (Poon and Maher,
1997) became the dominant thinking with the
development of CAD tools (Computer-Aided
Design). Evolutionary computation is concered with
avoiding the evolving of large numbers of
unsatisfactory solutions. In order to do so, a
subjective evaluation method is needed, but it is hard
to define. In this paper, a preliminary attempt on
evolutionary exploration based on subjective
evaluation is presented. Our aim is to build
connections between a logical system and subjective
space of design. A novel aesthetic evaluation model
embedded in generative design exploration
framework is introduced, and an Embroidery Design
System is built to prove the validation of this theory.
2 DESIGN EXPLORATION
For the purpose of creative design, especially in the
art domain, traditional evolutionary method does not
solve these ill-defined issues very well, which have
many uncertainties, imprecise descriptions and
subjective assessment criteria (Antonsson and
Sebastian, 2005). In this section, we introduce a
novel aesthetic evaluation model to build a bridge
between subjective and physical evaluation aspects,
and a heuristic exploration algorithm is presented for
evolutionary computation.
2.1 Aesthetic Evaluation Model
Comparing the traditional evolutionary fitness
functions which are mainly focused on searching for
optimization process or constraining satisfaction
process, exploration is more suitable for design
domain in order to satisfy designers’ potential
requirements guided by the vague assessing criteria.
Therefore, a multi-dimensional evaluation model is
introduced here. There are two kinds of evaluation
dimension: One is subjective dimension to express
different feeling and preferences in a designer’s
mind; another is the physical dimension to evaluate
the concrete design features.
The physical dimensions are linear in Cartesian
Coordinate, which start from origin and their
absolute values are increased with the amount from
their distances to the origin (equation 1). However,
the subjective dimension (equation 2) is quite
different. The value of a subjective dimension is
non-linear without the global extremum, but several
local extremums. In other words, the value of
subjective dimension is unstable which fluctuates
around a relative statistic point. So, there is no
absolute value but many relative values.
There is a multi-to-multi map between the two
kinds of dimension. The subjective dimension is
fluctuated and data-sensitive. Small changes from
physical dimension can bring big alternations in the
303
Cui J. and Xi Tang M..
GENERATIVE DESIGN EXPLORATION FRAMEWORK BASED ON SUBJECTIVE EVALUATION.
DOI: 10.5220/0003669003030306
In Proceedings of the International Conference on Evolutionary Computation Theory and Applications (ECTA-2011), pages 303-306
ISBN: 978-989-8425-83-6
Copyright
c
 2011 SCITEPRESS (Science and Technology Publications, Lda.)

subjective one, and vice versa (equation 3).
Dimension
subjective
= E
subjective
;
Dimension
1
, Dimension
2
,…, Dimension
m
= E
1
, E
2
,
…, E
m
, which m>0;
E
i
= γ* x
i
+ η, which m>0, γ, η∈R;
(1)
E
subjective
= Φ + α* sin(β*x + q), whichΦ,α,β,q∈R;
(2)
E
subjective
= 
E
1
E
2
…
E
m
 * 
a
1
a
2
…a
n
b
1
b
2
…b
n
… …… …
m
1
m
2
…m
n
 * Ψ,
which m, n ∈ R
+
;
(3)
In equation (3), Ψ is a fluctuated factor used to
control the sensitive trend; when Ψ is larger than 1, a
small change in the physical dimension will cause
big fluctuated changes in the subjective dimension;
When Ψ is smaller than 1, a small change in
subjective dimension will bring huge value changes
in the physical dimensions.
The matrix of 
a
1
a
2
…a
n
b
1
b
2
…b
n
… …… …
m
1
m
2
…m
n
 is the feature
matrix which can be specified by a designer for
pending or ongoing generating process.
2.2 Exploration Algorithm
The exploration process is an application
incorporating a human designer’s interference and
computational operation. For every generation,
designers need specify an evaluationary value based
on their subjective assessments. Then, the
evolutionary computational mechanism calculates
(3) to get every fitness value of every physical
dimension for changing the evolutionary step in the
next generation. Through this algorithm, we could
trace the influence of subjective fitness value to
physical fitness value for future research.
The algorithm is presented as follow:
Exploration Algorithm
Begin
1. To create the design generation based on random
possibility model;
2. To evaluate the generation by user and specify the
value of E
subjective
;
3. To calculate every fitness value of every physical
dimension E
i
of equation (3);
4. To decide whether change the feature matrix of
equation (3);
5. To compare the last evaluation with the current one. If
the last is larger than the current, then change the
developmental direction of the parameters, otherwise,
keep the same developmental direction;
6. To calculate the change pace by E
i
;
7. To get the new parameters based on E
i
and feature
matrix;
8. Io generate new results based on the new parameters;
9. If the result is satisfactory, then end the process, or, go
backto step 2.
END
3 EMBRODERY DESIGN
SYSTEM
Embroideries of Zhuang ethinic has long history and
folkoric tradition in Chinese Yunnan Province. The
design patterns have strong cultural characters on
shape and color, pursuing the simple and decorative
beauty with impressionistic and abstractive
aesthetics. In this section, a generative framework,
Embroidery Design System, makes use of generative
productive abilities and design exploration algorithm
to generate novel design solutions.
3.1 System Structure
There are three main parts in the system: User part,
Application part, and Evolution part. The User
part concentrates on the interactive function with
users, including the information input, information
output and visual interface; The Application part is
an aggregation of data reservoir and visualization
generator; The Evolution part focuses on the
evolutionary computation and design exploration.
In this system, users can type the data
information to specify their preferences through
valuing the parameters, as well as, operate
instructions to draw their favorite petal patterns
using a visual interface. Consequently, the patterns
and parameters are preserved in a Pattern Database
and an Infor Database, of which the Control.Info is
sent to next component by the message system and
the data is sent to Visualization Unit for visual
generating.
3.2 Result Analysis
The Embroidery Design System is programmed by
Visual Studio C++ 2003 and ACIS in Window XP
system. The interface technique is implemented by
XTreme ToolketPro 2008.
For the exploration algorithm, we extract the
feature matrix as follows:
ECTA 2011 - International Conference on Evolutionary Computation Theory and Applications
304


0.9 0.7 0.5 0.2 0 0 0.1 0 0.2 0.1 0.1 0.2
0.1 0.1 0 0.1 0 0 0.3 0 0.6 0.7 0.7 0.5
0 0.2 0.5 0.7 0 0 0.6 0 0.2 0.2 0.2 0.3

In the subjective dimension of Aesthetic Evaluation
Model, the subjective criteria can be subdivided into
balance, redundancy and harmony according to the
view of romanticist of aesthetics (Reich, 1993). So,
(3) they can be translated into equations (4)-(15), in
which the groups of {a1, a2, … , a12}, {b1, b2, … ,
b12}, {c1, c2, … , c12} come from the feature
matrix.
E
cate
g
or
y
= a
1
*E
balanc
e
+ b
1
* E
redundanc
y
+ c
1
* E
harmon
y
(4)
E
partproportion
= a
2
*E
balance
+ b
2
* E
redundancy
+ c
2
*
Eharmony
(5)
E
sy
mmetr
y
= a
3
*E
balanc
e
+ b
3
* E
redundanc
y
+ c
3
* E
harmon
y
(6)
E
relationship
= a
4
*E
balance
+ b
4
* E
redundancy
+ c
4
* E
harmony
(7)
E
p
etalst
y
l
e
= a
5
*E
balanc
e
+ b
5
* E
redundanc
y
+ c
5
* E
harmon
y
(8)
E
petalcolor
= a
6
*E
balance
+ b
6
* E
redundancy
+ c
6
* E
harmony
(9)
E
s
temst
y
l
e
= a
7
*E
balanc
e
+ b
7
* E
redundanc
y
+ c
7
* E
harmon
y
(10)
E
stemcolor
= a
8
*E
balance
+ b
8
* E
redundancy
+ c
8
* E
harmony
(11)
E
p
etalnum
= a
9
*E
balanc
e
+ b
9
* E
redundanc
y
+ c
9
* E
harmon
y
(12)
E
patternheight
= a
10
*E
balance
+ b
10
* E
redundancy
+ c
10
* E
harmony
(13)
E
p
atternwidth
= a
1
1
*E
balanc
e
+ b
1
1
* E
redundanc
y
+ c
1
1
* E
harmon
y
(14)
E
partnumber
= a
12
*E
balance
+ b
12
* E
redundancy
+ c
12
* E
harmony
(15)
During the generation process, a user can only
specify these three subjective evaluation parameters
from the visual interface (Figure 1). There is no need
to interpret the meaning of the subjective criteria
(‘balance’, ‘redundancy’ and ‘harmony’), as
different users with different design preferences
must have different understanding about them. So,
these three subjective values are just as the standards
to weight the physical fitness value for the next
generation.
(a) (b)
Figure 1: The snapshot of interface. (a) The system panel,
(b) The visual interface.
In the system, users can use the digital system to
explore their favorite embroidery patterns using the
exploration algorithm mentioned in section 2.2, the
collecting data is in Figure 2.
(a)
(b)
(c)
(d)
Figure 2: The fitness value in evolutionary process.
In Figure 2, we can see that at the generation 8
and generation 23, the values (Figure2.a) are
comparatively centralized, and they can become
emanative after that. Meanwhile, the values of three
subjective criteria are all above 50 and then sharply
down below 20. It indicates that during the two
phrases, the results from evolution generations are
onverging to a fixed style, and the fitness values of
subjective dimentions are reaching at local
maximum during the evolutionary process, whilst
these are no user’s favorite designs. As a result the
values of subjective evaluation dimensions are not
0
20
40
60
80
100
120
1357911131517192123252729
category
symmetry
partnum
relationship
petalnum
stemstyle
0
20
40
60
80
100
1 3 5 7 9 11131517192123252729
balance
balance
0
20
40
60
80
1357911131517192123252729
redunancy
redunancy
0
20
40
60
80
100
1357911131517192123252729
harmony
harmony
GENERATIVE DESIGN EXPLORATION FRAMEWORK BASED ON SUBJECTIVE EVALUATION
305

kept at a relative high level. This situation of gradual
changes in the physical dimensions bringing the
sharp changes in the subjective dimension shows the
fluctuated and data-sensitive characters of the
subjective dimension evaluation method.
After generation 27, the three subjective fitness
values are all higher than 70 and are kept at this
level, whilst and the physical fitness values become
concentrated. It means that, in this case, the user
finds his/her favorite design patterns at the end of
evolutionary process.
It is noticed that through the exploration
algorithm, the user interactive operation is quite
different from the traditional IEA approaches. In our
system, every generation is created by the users’
fitness value from subjective dimension and the
computational fitness value from the physical
dimension. By analyzing the collected data, it is
found that there are some interrelationships between
the subjective space and the logical mechanism, and
this validates the feasibility of our aesthetic
evaluation model trying to connect the subjective
and objective world. There are some novel designs
generated from Embroidery Desigy system, in
Figure 3.
Figure 3: Some samples of creative design by EDS.
4 CONCLUSIONS
In this paper, The novel aesthetic evaluation model
and design exploration algorithm can represent the
subjective feeling and physical features of designers.
Through experiment, we can trace the process of
subjective and objective evolutionary evaluation and
analyze the feature of the relationship between them.
This is considered as an advanced interactive
evolutionary computation objective for the research
on AI in Design but so far little is tested with real
design examples.
However, at this preliminary state of our
research, some disadvantages need to be improved
in the future. Firstly, although the aesthetic
evaluation model can express the fitness value of
design objectively, but how to specify a feature
matrix in a dynamic way is our next destination of
the research. Secondly, based on the evaluation in
physical dimension, involvement of heuristic a
searching method for deciding the evolutionary
development is needed but would require more
studies, for achieveing more efficient and effective
exploration of design solutions.
ACKNOWLEDGEMENTS
This paper is sponsored by a Hong Kong UGC
research grant, the general research fund (project
number RGC 531209).
REFERENCES
Antonsson E. K., Sebastian, H. J., 2005. Fuzzy fitness
functions applied to engineering design problems,
European Journal of Operational Research, 166: 794-
811.
Poon, J., Maher, M. L., 1997. Co-evolution and emergence
in design, Artificial Intelligence in Engineering 11:
319-327.
Reich, T., 1993. "A model of aesthetic judgment in
design." Artificial Intelligence in Engineering 8(2):
141-153.
ECTA 2011 - International Conference on Evolutionary Computation Theory and Applications
306