Ant- and ant-colony-inspired ALife art is characterized by the artistic exploration of the emerging collective behavior of computational agents, developed using ants as a metaphor. We present a chronology that documents the emergence and history of such visual art, contextualize ant- and ant-colony-inspired art within generative art practices, and consider how it relates to other ALife art. We survey many of the algorithms that artists have used in this genre, address some of their aims, and explore the relationships between ant- and ant-colony-inspired art and research on ant and ant colony behavior.
Within the past decade a wide range of software-based art projects conducted by artist-scientists grounded in ALife, art, and visualization have emerged under the umbrella heading “ant- and ant-colony-inspired ALife art.” In this special issue contribution, we trace the development of this movement in the visual arts, discuss its relevance and stature within both art and ALife, and elaborate on some of the specifics that bind ant behavior research with ant- and ant-colony-inspired ALife art.
For the purpose of this article, we define ant- and ant-colony-inspired ALife art as the set of works where artists use ants as a metaphor for the creation of computational agents for artistic purposes so that they can explore the collective behavior that emerges through the interaction of such agents, independent of their biological plausibility. Thus we focus on the development, relevance, importance, and impact of artwork based on ant and ant colony behavior. The astounding variety and complexity of this behavior has captured the attention of artists and public alike, thanks in no small measure to popular surveys such as those by Hölldobler and Wilson  and Gordon .
Although this article deals exclusively with the visual arts, we note that our definition of ant- and ant-colony-inspired ALife art excludes many familiar examples of ant-themed visual artwork, including numerous examples found in Australian aboriginal art (see ) and the cast sculptures of ant tunnels by Tschinkel  or Forti , as well as kitsch art or craft art such as “fire ant art” (see www.extension.org/pages/32528/fire-ant-art).
In the next section we review the seminal research that created the conditions for the emergence of ant- and ant-colony-inspired ALife art. In the third section we present a chronological overview of relevant work in this genre. In the fourth section we discuss ant- and ant-colony-inspired ALife art within the context of algorithmic art practices, artistic motivation, and the relationship of such art to other forms of ALife art. In the fifth section we provide technical details of some of the algorithms ant- and ant-colony-inspired artists use and explore their connections to ant and ant colony behavioral research. In the final section we draw our conclusions.
In this section we recall some of the key research developments that preceded and helped foment the emergence of ant- and ant-colony-inspired ALife art.
Langton's ants  were mobile cellular automata maneuvering on a rectangular grid of black and white cells following two rules: At a white square, turn 90° right, flip the color of the square, move forward one unit; at a black square, turn 90° left, flip the color of the square, move forward one unit. There is a well-known 1991 video by Langton (see http://www.youtube.com/watch?v=w6XQQhCgq5c) showing various experiments demonstrating the self-organization and emergent structures that resulted when two (or more) virtual ants, or Vants, interact starting from various initial configurations.
Ramos et al. [44, 45] described ant colony simulation experiments where virtual ants roaming over grayscale images produced “cognitive maps” (i.e., pheromone fields) to represent the images. As an application of this pattern recognition research, they considered image and document retrieval.
This section provides a chronology of the ant- and ant-colony-inspired artworks we are aware of.
Tzafestas  developed a drawing tool, called Ant Brush, based on a behavioral model where members of an artificial ant society collect crumbs, in order to pick up and distribute color on a drawing canvas.
Barrass  exhibited at the Melbourne University Student Union's George Paton Gallery under the title “Laying Down a Path in Walking” and also gave an artist's presentation at the Second Iteration Conference. His objective was “to make (moving) images that integrate structure and change, showing the perpetual overhaul of structures in a process of continuous change” (refer to online supplementary materials at www.mitpressjournals.org/doi/suppl/10.1162.ARTL_a_00170).
Monmarché et al.  used four to six virtual ants roaming over a toroidal grid, which painted by depositing scent in the form of one color, while searching for scent of a different color. The abstract images that resulted were interactively evolved. Although the ants were still essentially mobile cellular automata, the genome was considerably more sophisticated than the simple lookup tables of Langton's Vants or Turk's Turmites. Monmarché's intent was to find a more expressive way to understand and explain some of the ant algorithms that were then being developed by ant colony optimization researchers. Monmarché was the first to use the term “ant painting” for the static images that he created.
Barrass  began exploring the artistic potential for modeling ant behavior using neural nets. One example of the videos he produced was included in the 2006 EvoMUSART art show and has also been published online .
Semet, Durand, and O'Reilly  combined image processing techniques with ant colony simulation methods to develop nonphotorealistic rendering methods. What distinguishes their work is the use of swarms of ants to yield representational, as opposed to abstract, imagery derived from photographs.
Greenfield  evolved ant paintings using a model with 8 to 12 ants, similar to that of Monmarché. However, the fitness measures were based on ant behavior statistics. To the best of our knowledge, this was the first evolutionary approach to ant paintings that performed autonomous evaluation of the paintings, that is, did not rely on the user for fitness evaluation.
Urbano [56, 57] considered an ant painting model where individual cells in the environment exuded virtual scent to attract ants. Because the cells ceased to exude scent when they were first visited by an ant, once all cells were visited and painted according to colors assigned to the ants, the ant painting was finished.
Greenfield [19, 20] further developed Urbano's model by having the ants also deposit virtual scents in the environment that could act as either attractants or repellents. Therefore, the approach introduced a more plausible biological model that combined the use of multiple pheromones with evolutionary algorithms.
Greenfield and Machado  used a grayscale version of an ant painting model as a software component in a broader “artists and critics” simulation. Greenfield  then developed a colorized version to further investigate the artistic potential of such ant paintings.
Fernandes , adapting a technique used previously in image processing whereby artificial ants developed pheromone patterns by roaming over grayscale images [41, 13], developed a nonphotorealistic system similar in spirit to that of Semet et al., which he called Pherographia .
Urbano  became the first to execute ant paintings faithful to a biological model. Appealing to research of Franks et al. [15, 16] on nest construction by the ant species T. albipennis, Urbano's virtual ants used a stochastic sand grain foraging, collecting, and depositing algorithm in order to construct circular walls defining their nest's outermost boundary.
Greenfield  modified an ant colony model proposed by Jones , originally used for visualizing the evolution and formation of plasma transport networks of the slime mold Physarum, to develop ant paintings, which he called “network transport overlays.”
Greenfield [24, 28] used the same nest building algorithm as Urbano , but with carefully chosen radii, centers, and sand grain color preferences, in order to create ant paintings consisting of symmetric patterns of circles.
Machado and Pereira  introduced Photogrowth, which, similarly to Semet et al. , created abstract nonphotorealistic ant paintings from color photographs, producing artworks characterized by the intertwinement of lines of varying width and transparency. The evolution of sets of “sense vectors” for the ants and the production of scalable vector graphics as output are two of the novel aspects of their approach.
Machado et al. [36, 38] modified their Photogrowth system by allowing its users to design fitness functions, which were then used to guide evolution. Following the work of Greenfield and Machado , the fitness functions were based on both behavioral and image features.
Fernandes et al.  further enhanced their Pherographia system so that it created abstract images from color photographs.
Greenfield  extended the Urbano nest building model used for artistic purposes so that nest “boundaries” could be specified using polar or Cartesian curves.
4 Art Theory Perspective
In reflecting upon the history of art and algorithmic procedures, Roman Verostko, one of the founding figures of the algorithmic art movement (see http://www.algorists.org) that emerged in the 1960s, recently noted, “The period from 1980 to the end of the century remains an extremely rich period yet to be reviewed and understood better” (Roman Verostko, pers. comm.). To elaborate further, when paraphrasing the artist Peter Weibel's thoughts on the algorithmic movement, in 2004 Verostko observed:
The algorithmic revolution lies behind us and nobody noticed it. That has made it all the more effective—there is no longer any area of social life that has not been touched by algorithms. Over the past 50 years, algorithmic decision-making processes have come very much to the fore as a result of the universal use of computers in all fields of cultural literacy—from architecture to music, from literature to the fine arts and from transport to management. The algorithmic revolution continues the sequencing technology that began with the development of the alphabet and has reached its temporary conclusion with the human genome project. No matter how imperceptible they may be, the changes this revolution has wrought are immense .
4.1 Artist Motivation
Despite this state of affairs, the use of agent-based simulation in ant- and ant-colony-inspired ALife art establishes that it is firmly embedded in the traditions and practices of algorithmic, computational, and generative art. The effect of incorporating the ALife challenge of synthesizing ant and ant colony behavior into artworks builds upon the art tradition of capturing and emphasizing performance and process over time. To expand upon this point, kinetic art notwithstanding, just as Duchamps captured the passage of time with his Nude Descending a Staircase and Pollock focused upon artistic processes in his drip paintings, (static) ant paintings capture ant and ant colony behavior over time while, perforce, ant-inspired videos and installations do the same. The swarm robotic paintings of Moura, which resulted from a collaboration with Ramos that built upon the ant-based projects mentioned above, have been widely exhibited.1 There is an almost hypnotic attraction to seeing videos of Pollock executing drip paintings or Moura's robots drawing, which no doubt helps contribute to their success. Such an attraction is also, at least in our minds, present in watching the unfolding work of a colony of virtual ants as they create emergent patterns.
We argue in favor of both the effectiveness and the untapped potential of ant- and ant-colony-inspired ALife art on the basis of its increasingly close ties to observational and experimental biological research into animal behavior. Our thesis is that understanding such behavior may well be the emerging theme of the twenty-first century. When coupled with the lure of capturing processes over time as discussed above, for some artists and theorists this becomes compelling .
Beyond the initial heralding of ALife art via (interactive) genetic algorithms that was garnered by Sims  and Todd and Latham  under the banner of evolutionary art, perhaps the greatest impact on the art world associated with ALife art can be attributed to Kac's transgenic GFP Bunny Alba (see http://www.ekac.org/gfpbunny.html)  and the Vacanti mouse . This attention was in response to societal concerns about research involving genetics, the genetic code, and the ramifications of DNA databases. Thanks to growing scientific interest and public debate about crowd intelligence and distributed computation with (small) mobile devices, coupled with the never ending quest to understand animal behavior, ant- and ant-colony-inspired ALife art hopefully will also eventually merit such attention.
4.2 ALife Art Context
There is no definitive taxonomy of ALife art, and for this reason we will only briefly consider the positioning of ant- and ant-colony-inspired ALife art within the broader context of ALife art. There are several interrelated disciplines that produce art works and artefacts overlapping with the use of ant and ant colony behavior that may confound or confuse the discussion about the significance and status of any one them within contemporary art in general, and ALife-inspired art in particular.
4.2.1 Swarm Art
Certainly ants qualify as swarms, but more often than not “swarm art” refers to art grounded in flocking, herding, or schooling behavior algorithms that build upon Reynolds' seminal set of rules from 1987  that were the basis of his Boids simulation (see http://www.red3d.com/cwr/boids/). This leads to an extreme divergence between the ant and ant colony art described in Section 3 and the swarm art inspired by flocking algorithms such as the three-dimensional version used by Jacobs et al. .
4.2.2 Agent-based Art
Here, we refer to works that trace their lineage to Annuziato's 2002 agent-based art installation Relazioni emergenti . These agent-based works have an ecological flavor, often focusing on niche formation or environment sculpting. For example, in view of its description, in this genre Driessens and Verstappen's 2005 E-volved Cultures  immediately comes to mind:
“E-volved Cultures” (2005–2011) is a software presentation in which an artificial landscape grows in real time. Virtual agents that leave visual traces in interaction with their environment generate the dynamic pixel-landscape. The colourful abstract animations arouse associations with landscapes, geological processes, cloud formations, fungal growth, organ tissues or satellite photos, but ultimately they still avoid any definitive identification .
4.2.3 Collective Robotic Art
As noted previously, Ramos' 2002 ant colony simulation experiments led directly to the physically embodied collective robotics paintings of Moura [42, 43]. We note that even though his robot controllers were hand-crafted, Moura promoted these paintings as “nonhuman” art. Therefore, from an ALife standpoint, what is of principal interest is the nature and manner of the controllers developed for the drawing robots. No collective behavior was evolved. This is also the case in the work of Greenfield , where evolving controllers for drawing robots was reduced to evolving executable programs, and the use of multiple robots was a side issue.
5 Fields Related to ALife Research
This section discusses technical details and implementation issues that have drawn from, or have a direct bearing upon, ALife research.
5.1 Neural Nets
To make videos and sonifications of simulated ant behavior , Barrass used one-dimensional Kohonen  self-organizing maps (see, e.g., http://www.ai-junkie.com/ann/som/som5.html). Neural mappings were developed by sensing the traces ants left in such a way that ants learned to respond to familiar patterns, influencing each other indirectly through a shared artefact they produce and inhabit. The input to the self-organizing map from the ant's “feeler” was a sampled pattern of ant markings, and the activated output node yielded the turning angle to use for updating the ant's position. Thanks to gradually fading ant trails, several clever and innovative ways of updating the self-organizing map, and judicious choices for drawing attributes, Barrass obtained an impressive and diverse collection of digital prints and videos. Barrass carefully stated:
I should note that the ant metaphor is only a starting point for the work, and a convenient way to think about the agents in the system. I am not making any serious comparison with the behaviour of living creatures [3, p. 61].
5.2.1 Monmarché, Greenfield, and Urbano—Primitive Pheromones
The genetic algorithms used for evolving ant genomes are straightforward. However, it should be noted that what is evolved in this context is actually a small collection of such genomes, so that the resulting ant “societies” can evolve into castes with different color preferences, thus promoting interaction and feedback. Monmarché uses an interactive genetic algorithm with 4 –6 ant genomes, while Greenfield uses an automated genetic algorithm with 8 –12 ant genomes. Greenfield's fitness functions are simple arithmetic expressions involving the number of time steps ants spent exploring (i.e., seeking their preferred scent) and the number of time steps ants spent exploiting (i.e., following their preferred scent). In this way ant paintings are evolved on the basis of collective ant behavior.
Urbano [56, 57] used a different setting. He treated each cell in the toroidal grid as an ant chemical emitter, continually emitting an ant attractant until it was first visited by an ant, at which point the cell was painted a different color than the original background gray and ceased to function as an emitter. The attractant diffused and evaporated. Using this method, an ant painting was finished once all the cells had been visited. The amount of chemical that was emitted and the evaporation and diffusion rates were user modifiable parameters.
5.2.2 Greenfield—Multiple Pheromones
While Urbano did experiment with an ant collision avoidance policy—when an ant senses another occupant in its cell, it jumps to a nearby cell—it was Greenfield [19, 20] who incorporated the use of an invisible evaporating and diffusing pheromone emitted by the ants themselves to first address this problem. When the attractant emitted by the cells was too weak, the ants used the gradient of its pheromone to try and avoid regions of the environment that had already been explored. To leverage this behavior for artistic purposes, when in avoidance mode ants deposited a new color that was blended and diffused into the painting. Figure 1c shows a representative ant painting produced using this technique. An interesting side effect was the rotational patterns of ant movement that emerged as a result. Although it was caused by avoidance rather than attraction, it raises the question of whether ant colony simulations can give rise to the so-called death circles, or ant mills, that occur when excessive ant attractant pheromone causes ants to circle endlessly following one another until they die . Another innovation of this work was to initially cluster two colonies of ants (using two different color schemes), as opposed to sprinkling large numbers of ants randomly throughout the environment as was normally done. The rationale for sprinkling is that the randomness will disappear during a transient phase. Greenfield used an evolutionary algorithm to consider how the initial placements of these two colonies could guide the formation of the paintings.
5.2.3 Semet, Fernandes, and Machado—Complex Environments
Ant-inspired models used in nonphotorealistic rendering use the pixels of a source image to serve as the environment; hence they often furnish a rich, complex environment. We consider three nonphotorealistic rendering systems that are based on simulating ant behavior.
Semet et al.  use an interactive genetic algorithm to let users control the composition and style of their nonphotorealistic ant paintings. Ants roaming over the environment deposit marks on a separate virtual output canvas. Many of the parameters under user control affect ant drawing attributes. In addition to sensing the RGB color of each pixel, ants are also able to sense three preprocessed quantities associated to each pixel: luminance, norm, and orientation of the gradient map derived from the luminance map (computed using 3 × 3 Sobel filters), and from the salience of the pixels, which indicates they are part of key features or regions of interest. Because ants sense the local environment without changing it, and their actions depend only on what they encounter, a colony of several hundred ants is simulated by having one ant at a time roam over the environment for a fixed number of time steps. Ants shift among executing a small number of hand-crafted sensing and drawing algorithms designed for mark making purposes.
In the system Photogrowth by Machado and Pereira , ants also sense an image environment while creating a nonphotorealistic rendering by laying down “ink” on an output canvas. They adopt a metaphor where brightness is perceived as food. The luminance of an area of the image environment determines the amount of food available at that point. Each ant has an initial energy, which decays over time, but ants can gain energy by traveling through illuminated areas of the image environment. Energy governs the birth-and-death cycle. The energy of an ant also determines the width of the trail drawn on the output canvas. Thus variation in energy results in trails of varying width. The consumption of energy affects the image environment, implying that, unlike those of Semet et al., the ants must be simulated in parallel because the behavior of one ant affects the behavior of others. Each ant uses ten sense vectors that determine ten positions relative to the current position of the ant, which in turn determine ten brightness readings obtained from the image. A weighted sum of these readings is scaled according to the ant's velocity to determine how to update the ant's position. Perlin noise is added to this calculation to simulate the uncertainty in the sensory readings and movement.
The parameters governing the behavior of the ant, including the direction and magnitude of the sense vectors as well as the corresponding weights, are encoded in the genotype and subject to evolution. Subsequently, Machado et al. [36, 38] abandoned using the interactive evolutionary algorithm to set parameters for their system in favor of having users design their own fitness functions, by having them weight a linear combination of components that amalgamated image statistics, image qualities, and ant behaviors. Figure 1d shows art work produced using this system (refer to online supplementary materials for additional examples).
We wish to point out that there are two options for implementing agents, thought of as ants, roaming on a two-dimensional gridded environment of cells and that these options may result in significant visual differences among outcomes. Either agents can jump from cell to cell, thus effectively acting as mobile automata, or they can move more smoothly through the environment by maintaining position, orientation, and velocity and then relying on rounding to sense and modify grid cells. In the discrete case the environment is realized as a lattice, while in the continuous case it is realized as a surface. Greenfield [19, 20], Fernandes , and Semet et al.  are examples of the discrete method, while Greenfield [25, 24] and Machado et al. [36, 38] use the continuous approach.
5.3 Remote Sensing
Jones  modeled the evolution and formation of plasma transport networks of the slime mold Physarum using an agent-based simulation whose agents, thought of as virtual ants, use remote sensing, that is, sense the environment several units distant from their current position. Here, a plasma transport network refers to the flow of protoplasmic sol through a gel matrix of fibers. Although the patterned images Jones obtained often resemble the reaction-diffusion patterns obtained using the well-known algorithmic methods of Turk , Witkin and Kass , or Young , they are, in fact, visualizations of the pheromone distributions of evaporating and diffusing pheromone trails laid down by virtual ants. Greenfield  composited colorized versions of these transport networks to create ant paintings that he called “abstract overlays.”
5.4 Foraging Behavior
The most recent use of ant behavior for ant- and ant-colony-inspired ALife art draws its inspiration directly from research biology. As it turns out, both examples involve ant foraging behavior.
5.4.1 Urbano—Sand Grain Foraging
Interestingly, this algorithm is remarkably similar to a “crumb laying” algorithm for artificial ants that was analyzed by Tzafestas  while investigating a robot foraging task. When ants are assigned different preferences for sand grain colors, Urbano called ants executing the algorithm given above “sand artists”. Both Urbano  and Greenfield [24, 28] have used variants of this sand grain foraging model to create ant paintings. Figure 1e shows a circle pattern design of Greenfield that was rendered using this method.
5.4.2 Greenfield—Seed Foraging
The desert ant species Pogonomyrmex barbatus does not use pheromone trails when foraging for seeds. It is understood, however, that the nest mound of these ants has a hydrocarbon profile that helps returning ants find the nest entrance , and it is known that in order to initiate seed foraging, each morning a number of patroller ants lay down several short pheromone trails on the nest mound to indicate directions foragers should use for that day's foraging. For their first foraging trip of the day, foragers proceed in one of the directions indicated by the patrollers for a considerable distance away from the nest, at which point they begin searching for a seed. When one is found (nearly all trips are successful), the forager returns directly to the nest, and then, for each subsequent trip, the forager returns to approximately the same location where the first seed was found to resume the search for seeds. Based on experimental evidence obtained by Gordon  that seed foraging is sustained on the basis of a high enough rate of ant-ant interactions occurring between successful returning foragers and queued idle foragers waiting at the nest entrance, and working from parameter values estimated from primary source materials, Greenfield  simulated the foraging behavior of a colony of these ants over a five-hour period. He then followed up by creating ant paintings based on the simulation of all the first foraging trips of the day .
We have endeavored to document the emergence and history of ant- and ant-colony-inspired ALife art and to understand how such art works relate to other ALife art genres. By examining the technical aspects of artists' projects, we have demonstrated how biological research and ant and ant colony simulation efforts have informed ant- and ant-colony-inspired ALife art.
The recent works of Urbano  and Greenfield  draw their inspiration from biological models of nest construction and seed foraging behavior. As such, their work can be seen, simultaneously, as artistic endeavors, as information visualization projects, and as model validation efforts contributing to a better understanding of ant behavior.
Along with such efforts grounded in biological models, we observe works where ant and ant colony behaviors are used as a powerful metaphor for the construction of a complex system. This departure from biological plausibility is openly and explicitly acknowledged by the artists, as evidenced from the early works of Barrass —which focuses on the visualization of simulated ant behavior—through the recent works of Machado et al. [36, 38], where an evolutionary computation system evolves the behavior of virtual ants, yielding nonphotorealistic renderings of input images in accordance to fitness functions designed by end users.
The body of work surveyed in this article reveals one of the most valuable characteristics of ant and ant colony research: the ability to provoke and inspire, making researchers push boundaries in an endless quest to understand, mimic, or match the untamed beauty of nature by all means available.
We are grateful to Alan Dorin for his advice and guidance during the preparation of this article. We thank Tim Barrass for providing us with access to his thesis as well as supplementary material. We thank Roman Verostko for helpful discussions about the origins and impact of algorithmic art.
At about the same time Ramos was using ants for pattern recognition, he initiated a project, “Machines of Collective Conscience,” and began collaborating with Moura. In one relevant project they used drawing robots to “paint” over large photographs of iconic artworks and figures (e.g., Picasso's Guernica and Marilyn Monroe); in the other, swarms of robots painted on a blank canvas. These large swarm drawings were exhibited internationally.
Department of Mathematics and Computer Science, University of Richmond, Richmond, Virginia 23173. E-mail: firstname.lastname@example.org
CISUC, Department of Informatics Engineering, University of Coimbra, Polo II of the University of Coimbra, 3030 Coimbra, Portugal. E-mail: email@example.com