Final Paper

bigskymanAI and Robotics

Oct 24, 2013 (4 years and 7 months ago)



Zach Joyce

Dr. Johnson

Evolutionary Computing

April 28, 2013

Human Based Genetic Algorithm to Create Ignatian Station Logo

Genetic algorithms have been shown to
produce successful solutions for
problems with
difficult or extremely large solution
possibilities. These genetic algorithms are, however, not
without their shortcomings. When topics arise that incorporate concepts that are difficult to
quantize, represent, or evaluate computationally, basic genetic algorithms fail to be of sufficient
lp. These problems can be overcome with the help of human involvement in genetic
algorithms, especially through Interactive Genetic Algorithms (IGA) and Human Based Genetic
Algorithms (HBGA). Of the primary genetic operators in a genetic algorithm, IGAs
use human
agents to evaluate fitness whereas HBGAs use

human agents for all operators
(population/selection, crossover/mutation, and fitness).

Human involvement in genetic algorithms is both interesting and useful because it allows
for the expansion of th
e field of GA’s to cover problems not easily quantized. Such problems
include the development of natural language, art/music, architecture, and design (as well as many
others). These areas are all difficult to evaluate using a computational fitness funct
ion, because
fitness levels are difficult to generate without greatly reducing the complexity of problems in
these areas. For areas such as language and design, genetic operators such as crossover and
mutation are also difficult to encode, and so HBGA wou
ld solve these problems through
increased human involvement. Another interesting feature of IGAs and HBGAs is the ability to
utilize humans’ vast and varying background knowledge, and psychological processing function

(intuition, preference, sensation, et


. In this way, human involvement allows
genetic algorithms to address and solve problems never before thought possible with pure
computational methods.

Using human involvement to increase the functionality of genetic algorithms started in
he early 1990s, but fully implemented HBGAs were not presented until 2001. Work with IGAs
was first introduced by Richard Dawkins while discussing computational evolvability. They
were brought to light by Craig Caldwell and Victor Johnson in 1991, who we
re able to develop a
facial recognition system that learned using a genetic algorithm with human evaluators of fitness.
Japanese researcher Hideyuki Takagi was also a big influence of the world of IGAs, doing a lot
of work in the development and synthesis

of different algorithm strategies and implementations.
HBGAs were conceived by Alex Kosorukoff, building off the idea of outsourcing other genetic
operators to human agents. Kosorukoff is the father of HBGA and does most of the major work
in that area.

For my project, I was really inspired by

and Nickerson



HBGA to evolve designs for chil
dren’s chairs. This paper got m
e introduced to the idea of
HBGA and I wanted to explore what other areas could be advanced by using humans in the
genetic algorithm process. For my project, I decided to recreate Yu’s experiment, but use the
HBGA to design/evolve something other than cha
irs. Since I would be performing this
experiment with the help of my peers, I wanted to target the problem towards their interests. I
ultimately decided on evolving a logo concept for the new convenience store on campus, the
Ignatian Station.

The conv
enience store is a relatively new addition to Gallagher, and so does not yet have
a logo, but is
well recognized enough that anyone I asked to help would be motivated by their

Xavier pride to produce good designs. The problem of designing a logo fits well

into the
category of HBGA. A logo has many different aspects that appeal to people, some which can be
quantized (colors, size, shape) and many other which cannot (familiarity, is it
recognizable/memorable, etc). These aspects are also difficult to put i
nto bit strings that a
computer could reorganize
or evaluate
effectively. By using an HBGA to develop the logo, it
was my goal to use the genetic algorithm process to produce better and better logos with the help
of Xavier students for both the design and

evaluation/fitness process.

Before I could carry out the experiment, I had to decide on a number of parameters and
how I was going to represent the genetic operators throughout the process.
Each generation had
a population of 12 individuals. For each i
ndividual in the population, since I was using HBGA,
each logo was a concept sketch for a logo idea saved
as a picture (JP
G). Human agents were
involved in both the design process to create the population (by drawing the pictures), and were
also involved
in the fitness evaluation. For fitness, each logo was evaluated by 15 human agents
and graded on 5 of the t
op qualities of logos (“Characteristics”, np
) using a 7 point Likert scale.
The average scores for each category were calculated (for each logo) an
d then combined using a
weighted average in order to get a fi
tness score for an individual.

The selection process was then carried out using the tournament style, where four
individuals were picked and the two fittest assigned as a parent pair for the nex
t generation. This
was done for eight pairs of parents, which were presented to human agents as inspiration for the
creation of the next generation of logos.
Crossover and mutation, the usual genetic operators,
were left to the discretion of the human ag
ents producing each logo. Crossover was meant to be
encapsulated by presenting the two parent logos as inspiration for the next drawing, and mutation
can be thoug
ht of as any new elements introduced/replaced by the human agents.
The remaining

four spots
in the successive generation were filled th
rough elitism (the top two scoring logos
advanced) and new genetic introduction (human agents were prompted without inspiration).
Because of time restraints, the experiment was

for three generations, the same

as the Yu

For the actual experiment, I employed the help of 56 X
vier students to create and
evaluate different logo designs over three generations. For ease of use a
nd portability, most of
the human involvement was done using an iPad
In order to collect and create designs, human
agents were given an explanation of the project, the convenience store, and their role in the
process. They were s
hown a prompt

(see Figure 1)
and then asked to create a concept sketch of
their logo idea on the

app for iPad. Human agents were asked to create a concept
sketch in order to compensate for differences in artistic abilities. After completion, the drawings
re saved and uploaded
to be included in the fitness evaluation survey

Figure 1.

Prompt shown to human participants prior to
designing logo.


Figure 2.

Screenshot of the survey presented to
human evaluators, including logo picture and
scoring scale.

The fitness evaluation was
completed using the

app fo
iPad. Each logo in the population for a
given generation was included in a single
survey given to human evaluators. Each
picture was displayed with the same prompt
to view the image and evaluate it on five
criteria: 1) Appropriateness and

to its cause; 2) Originality and
distinct from existing logos; 3) Memorable;
4) Versatility and Adaptability; 5)
ness and Classic Look of the

Each attribute was scored on a 7
point Likert scale, as descri
bed above (see
Figure 2
). In ter
ms of generating a fitness score, the mean scores for each attribute were
combined using a weighted average with weights being (in order as listed above) 40%, 20%,
15%, 15%, 10%. The weights were chosen arbitrarily based on the relative importance of all
quality attributes (“Characteristics”, np

After fitness was calculated, 8 pairs of parents were determined using tournament
selection. These 8 pairs of parents then served as inspiration for the next generation of
individuals, with 2 logos being carried over through elitism, and 2 more being adde
d each
generation as brand new concepts.
This whole process was completed (creation of designs and

Figure 3.
Recombination process including two parent logos (left) that served as

inspiration for the
child logo in the next generation (right).

Figure 4.

Examples of logo designs produced. The t
op rated design (right) features a comb and toothbrush in
the shape of an “X” and the lowest rated design (left) is plain and features only the initials “I.S.”.

fitness evaluation) for three generations, ultimately producing a top design from the whole

The results of this experiment showed a wide range of values

(see Table 1)
. The total
average fitness increased over the course of the experiment, going from 4.39 (out of 7) after the
first generation, to 4.59 after the third and final generation. Looking at the
different parts of the
total score, the Distinct and Memorable ca
tegories each increased from the first to the last
generation. The other three areas, however, each either remained constant over time or



decreased slightly. These differences tend to suggest that human evaluators valued originality
and memorability more

than the other qualities, and so evolved the logos based on these
regardless of how the attributes were weighted.

It is interesting to note the trend of fitness when the second generation is considered. The
second generation had the highest rankings in
terms of almost all categories, as well as total
overall average fitness of all individuals. On a more individual basis, generation two also had
the most logos scoring above 5 (5 logos did so) than both generation one (2 logos) and
generation three (1 log
The reasons for this jump are unknown, although possible reasons
could include the addition of good new genetic material in this generation, or perhaps a favorable
fitness evaluation group. The decrease in fitness level from generation two to three i
s against the
goal of the experiment to show human collaboration, however it is unknown whether the
decrease is a general trend or a temporary setback without running the experiment for more
generations (not possible here because of time constraints).

and Applicable

Distinct and


Versatile and

Timeless and


Gen 1







Gen 2







Gen 3







Overall, I thought that the experiment was moderately successful, however I would have
liked to carried out more generations in order to determine a general trend for the logos.
Regardless of the number of generations, the top design after the third gener
ation was well
received, scoring above 5 in all evaluated criteria. I think that the process of implementing the
HBGA went fairly well, although there are some improvements I could make if I were to repeat
the experiment. In Yu’s paper, the experiment wa
s performed entirely online (through Amazon
Table 1
. Average fitness scores for each generation,
categorized under each quality attribute as
well as total fitness.


Turk), which called into question the motivation of its participants. I tried to improve that by
requesting involvement personally, and by picking a topic that most of my audience (Xavier
students) would care ab
out. I even noticed comments during the design process that some
participants wanted to create the best design, thus serving as self driven motivation to advance
the HBGA process.

Even though fitness was a set standard, I think that optimizing the
parameters by which
each logo is judged could produce more effective results. Perhaps even a different fitness
evaluation could be performed, one based on comparison rankings instead of individualized
scoring. Another possible set back with the fitness e
valuation was not adjusting for different
peoples’ rating scales. Even though each criterion was judged from 1
7, different evaluators
(especially across different generations) may have judged the values to mean different things.

Statistically adjusting
for this discrepancy would be something to add for future experiments.

Another way to perhaps produce better improvements in the designs would be to increase the
number of generations, and human agents in the process.
This is one area where the using of
Amazon Turk was better than my personalized method of human involvement.

Regardless of the shortcomings, however, the HBGA produced quality logo designs
through human collaboration. These ideas will be show to the Ignatian Station staff for

of use for their store. Throughout my time working on this project, I enjoyed
watching the algorithm run its course, as separate individuals all worked together to improve
designs towards a common goal.



Breukelaer, R., Emmerich, M., and T.
Bäch, “On Interactive Evolution Strategies,”
of Evolutionary Computing, International Conference on,
pp. 530
541, 2006.

Cheng, C. D. and A. Kosorukoff. “Interactive One
Max Problem Allows to Compare the
Performance of Interactive and Human
ed Genetic Algorithms,”
Genetic and
Evolutionary Computation (GECCO)
, pp 983
994, 2004.

“Characteristics of Great Logo Design.”
Mansueto Ventures LLC.

18 Mar 2013. Web.
2010. <

s, R. “The evolution of evolvability”,
Artificial Life

(Ed. C.G. Langton), Addison
Wesley, pp. 201
220, 1989.

Kosorukoff, A.
Human Based Genetic Algorithm (HBGA)
. 2000

Kosorukoff, A. "Human based genetic algorithm,"

Systems, Man,

and Cybernetics, 2001 IEEE
International Conference on

, vol.5, pp.3464
3469, 2001.

Saito, Y. and T. Itoh, “
e: A Visual Music Recommendation System featuring
Interactive Evolutionary Computing,”
Visual Information Communication (VINCI) 2011

Takagi, H. “Interactive Evolutionary Computing

Cooperation of computational intelligence

human KANSEI,”
Soft Computing and Information, 1998 Intelligent Systems
International Conference on,
pp. 41
50, 1998.

Yu, L. and J.V. Nickerson, “Cooks or Cobbl
ers? Crowd Creativity through Combination,”
Computer Human Interaction (CHI 2011)
, 2011.