Painting Using Genetic Algorithm with Aesthetic Evaluation of Visual Quality - Technologies and Applications of Artificial Intelligence - page 125

an interactive fitness evaluation to address this issue. That is, human users act
like the fitness function to assign fitness values or to select parents and survivors
for EAs. This manner is useful since it combines the search capability of EAs
and aesthetic taste of human. However, EAs usually require thousands or even
million times of fitness evaluation, which makes interactive EAs impractical due
to the fatigue and decreasing sensitivity of human beings after long-term eval-
uation. The interactive EAs for creativity, therefore, have to compromise with
small population and short run. Some studies attempt to avoid user's fatigue.
For example, Machwe and Parmee [12] used meta-feature clustering to learn and
predict user's judgment. Llor et al. [11] applied support vector machine to syn-
thesize the subjective fitness function. Kowaliw et al. [7] developed the EvoEco
system to automatically detect and emphasize creative designs. A potential issue
at these methods is that users may need to spend more time in fitness evaluation
once the prediction results are wrong and needed to be corrected. Some research
works on design of fitness functions that can emulate human aesthetic preference
by using machine learning technologies such as neural networks [1] and coevo-
lutionary algorithm [5]. Recently, Liu and Ting [9,10] proposed using objective
evaluation metrics instead of human feedback to address this issue. They de-
signed the fitness function based on music theory to evaluate the compositions
and achieved satisfactory results.
This study aims to generate paintings by EA with aesthetic evaluation. Specif-
ically, we develop a genetic algorithm (GA) and propose using aesthetic vi-
sual quality of paintings for fitness evaluation. In performance assessment, we
implement the EvoEco system, an EA-based image generator, and replace its
interactive EA with the proposed method. The resultant paintings show the
effectiveness of the proposed GA with aesthetic evaluation. The remainder of
this paper is organized as follows. Section 2 introduces the EvoEco system that
uses an interactive EA to generate images. Section 3 presents our proposed GA
and describes the structure of chromosomes and the fitness function based on
aesthetic visual quality. Section 4 presents and compares the generated paintings.
Finally, Section 5 draws the conclusions of this study.
2 Creativity and the EvoEco System
Regarding creativity, Dorin and Korb claimed that “Creativity is the introduc-
tion and use of a framework that has a relatively high probability of producing
representations of patterns that can arise only with a smaller probability in
previously existing frameworks” [3]. According to this definition, they built the
EvoEco system to generate images. EvoEco is a multi-agent ecosystemic platform
based on interactive EA to evolve into generative art. The system consists of two
parts: the ecosystemic (generative) stage transforms a genome into a phenotype
(image), and the evolutionary stage selects and evolves chromosomes.
Following the EA framework, an image is represented as a chromosome in
EvoEco. A chromosome is composed of k agents and each agent is in charge
of painting during its lifetime. The evolutionary process of EvoEco to generate
images is as follows: