Evolving Families of Shapes

Filipe Assunção, João Correia, Pedro Martins, Penousal MachadoIJCAI 2015

Abstract

Traditional evolutionary art systems typically focus on optimizing a single phenotype to achieve a high aesthetic score. This research proposes a shift toward evolving "families" of related shapes, mimicking the way human artists develop cohesive styles or "periods." By evolving Context-Free Design Grammars (CFDGs) using Graph-Based Genetic Programming, the system generates sets of images that are stylistically consistent yet individually diverse. A multi-objective fitness function incorporating mean quality, consistency, and a target diversity level is employed to guide the evolution of these visual languages.

Introduction

In evolutionary computation applied to the arts, the goal is often to find the "perfect" image. However, artistic value often stems from the context of a larger body of work. A single painting by Frank Lloyd Wright or a piece from Picasso’s "Blue Period" is understood within the framework of the artist's broader visual language. This paper addresses the challenge of mathematically defining and evolving such a "family" of shapes, avoiding the extremes of monotonous repetition and chaotic unrelatedness.

Methodology

The core of the system is the evolution of Context-Free Design Grammars (CFDGs) rather than static pixel arrays.

Genotype and Phenotype

The genotype is represented as a set of production rules within a grammar. When this grammar is executed with different random seeds, it produces a variety of phenotypes—individual images that share the same underlying structural logic.

A single phenotype generated from an evolved Context-Free Design Grammar.

Fitness Evaluation

To evaluate a grammar, the system samples a set of images ($S$) and applies a three-part fitness function:

  1. Mean Quality: The average aesthetic score based on chromatic measures, specifically evaluating color distribution (Bell Curve) and relationships between adjacent colors (Contrasting Colors).
  2. Consistency: The standard deviation of quality across the set; lower variance indicates a more reliable grammar.
  3. Diversity: The average similarity between pairs in the set, rewarded when it approaches a target value ($\mu$).

Results

The researchers explored the impact of the target similarity parameter ($\mu$) on the resulting families. They found that a "sweet spot" exists around $\mu = 0.5$, where the generated sets are diverse enough to remain interesting but consistent enough to be recognized as belonging to the same family.

Evolved families of shapes showing stylistic consistency and internal diversity across different grammar types.

As shown in the results, the system successfully evolves grammars that produce distinct "species" of art, ranging from circular compositions to complex spiral structures. High similarity ($\mu = 0.7$) led to monotonous families, while low similarity ($\mu = 0.1$) caused the family identity to collapse as images lost visual consistency.

Discussion and Insights

The "context effect" is a key takeaway: an evolved population is often more aesthetically compelling as a whole than any single member. By evolving the grammar (the "language") rather than the pixels, the system learns to generate infinite variations within a specific stylistic constraint. This approach moves evolutionary art closer to the human creative process of developing a signature style.

Conclusion and Future Work

While the system effectively evolves stylistic families, it currently relies on hard-coded aesthetic measures, which do not fully capture the complexity of human artistic judgment. Future research could integrate interactive evolution to define quality and diversity more subjectively. Additionally, moving beyond abstract shapes to more complex representational forms remains a significant challenge for grammar-based evolutionary systems.

Summary updated

BibTeX
@inproceedings{assuncao2015evolving,
  author = {Filipe Assunção and João Correia and Pedro Martins and Penousal Machado},
  title = {{Evolving Families of Shapes}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}

Loading related content...

Figure