Aesthetics of Adaptive Behaviors in Agent-based Art Sofian Audry Concordia University, Faculty of Fine Arts, Montréal, Canada info@sofianaudry.com Abstract This paper focuses on a particular facet of this broader work: agent-based adaptive computational artistic installa- Since the post-war era, a number of artists have been explor- tions. It argues for an aesthetics of adaptive agents rooted in ing the use of embodied, artificial agents, in parallel to sci- the distinctive way behavior morphologies unfold in time. entific research associated to Computer Science in domains Among the significantly vast literature that exists in such as Cybernetics, Artificial Intelligence and Artificial media theory, art history and STS on the topic of arti- Life. While notions of adaptation and learning have been ficial agency and machinic life in art and science, most an extremely important component of that research, artists studies focus on related concepts such as embodiment/ and media theorists seem to have focused on the concept of situatedness [23, 6], autonomy [5] and emergence [4, 29], emergence. Whereas emergence offers a rich ground for art- while very few directly address questions of adaptivity and making, adaptation is an equally important, yet complemen- machine learning. tary dimension of it. In an effort to re-position adaptive sys- I hereby wish to fill this gap by proposing an aesthetics tems within the theoretical and practical field of agent-based of adaptive agent-based installations. My main objective is artworks, an aesthetics of computationally adaptive artistic to provide a description of the experiential mechanisms that installations is proposed in this article. To do so, I examine are made possible by adaptive behaviors in media installa- (1) the historical context surrounding adaptive systems; (2) tions by connecting the dots between the scientific perspec- its relationship with the concept of emergence; and (3) the tives over such systems and the aesthetic effects they afford. aesthetic potential of Machine Learning algorithms by ex- While artistic media installations cannot be separated from amining their intrinsic characteristics. An aesthetic frame- their visual and audio qualities, the focus of my analysis is work based on the morphological aspects of the temporal on their processual dimension. unfolding of agent behaviors is offered as a tool to compre- hend both adaptive and non-adaptive behaviors in works of This paper is concerned with providing conceptual tools art. to support reflection and creation by artists and researchers engaging with adaptive systems. To contextualize my re- search, I first present an overview of the history of adapta- Keywords tion from the 1950s onwards, focusing mainly on cybernet- Adaptive Systems, Aesthetics of Behavior, Agent-based ics, artificial life, and machine learning, showing their im- Art, Artificial Life, Cybernetics, Machine Learning, Media pact on new forms of art. Building upon Cariani’s catego- Art Installations. rizations of adaptive and emergent systems, Penny’s “aes- thetics of behavior” and Xenakis’ theory of morphological Introduction evolution, and looking at specific considerations surround- Since the 1960s, media artists have been creating bodies ing Machine Learning technologies, an aesthetic frame- of work using and/or inspired by computer technologies. work is put forth to understand the evolution of behaviors In this article, I am interested in a specific branch of artis- through time. tic works that make use of artificial agents, which include works such as Nicholas Schöffer’s cybernetics sculpture Historical Context CYSP1 (1956), Ken Rinaldo’s artificial life installation Au- topoiesis (2000) and Yves Amu Klein’s living sculpture History is imbued with a fascination for human-fabricated series. Artist and media theorist Simon Penny calls these life, from Al-Jazari’s 13th century’s moving peacocks to kinds of work “embodied cultural agents” or “agents as art- Jacques Vaucanson’s digesting duck (1739). A change in works” and integrates them within the larger framework of paradigm operated in the post-war era with the advent of an “aesthetic of behavior”, a “new aesthetic field opened up computers which, contrary to mechanical automata, are by the possibility of cultural interaction with machine sys- uniquely powerful in both their speed and programmatic tems” [24]. These works are distinct from so-called gen- capacity. But while often seen as fixed, logic-based sys- erative art which uses computer algorithms to produce sta- tems, an important strand of research in computer science bilized morphologies such as images and sound: their aes- rather focuses on their malleable, organic properties, ap- thetics is about the real-time performance of a program as it proaching them as adaptive, self-organized, statistics-based unfolds in real-time in the world through a situated artificial devices. This section offers an overview of this research body. while examining its role in contemporary media art. 2 Proceedings of the 22nd International Symposium on Electronic Art ISEA2016 Hong Kong. Aesthetics of Adaptive Behaviors in Agent-based Art. Sofian Audry Cybernetics, Perceptrons and Classic AI objects, but rather new information, embodied in the cre- ation of works of art [8]. The first conceptions of adaptivity in organisms can be found in the work of early, so-called “first-order”, cyber- neticians. Norbert Wiener’s notion of control in Cybernet- Artificial Life ics systems is closely linked to the concept of teleological At the beginning of the 1980s, classic approaches in AI or negative feedback. In a system that displays such neg- were still dominating, showing no interest in any form of ative feedback control, the difference between the goal of biologically-based computation such as genetic algorithms the system and its current outputs is sent back to the inputs, and neural computation. Nonetheless, two strands of re- allowing the system to correct its course; in other words, to search would come to life in that era, challenging the statu constantly adapt to small changes in its environment. [32] quo: Artificial Life and Machine Learning. A key and related concept is that of homeostasis, referring In the 1970s, chaos theory and complex system theory to the ability of living systems to maintain stability within had revealed how highly non-linear systems often display an unstable milieu using self-regulation. [3] emergent properties, that is, unpredictable behavior as the Building upon both cybernetician models of the brain [3, result of simple interactions between a large number of en- 19] and psychologist Donald O. Hebb’s theory of self- tities. Emergence directly challenges the distinction be- assembling neurons [14], Frank Rosenblatt proposed in the tween human and machine: starting from simple rules, we late 1950s one of the first adaptive connectionist devices, can simulate complex and unpredictable behavior on the the perceptron [25], a simplified model of a human neural computer. This idea is core to the early 1980s apparition of network that maps a set of binary data (input neurons) to Artificial Life (ALife), a synthetic approach to biology that a binary output (output neuron) using a layer of paramet- seeks to create “life-like behaviors”. This new “biology ric values called weights (representing the synapses) which of possible life”, directly influenced by Cybernetics, sup- are initialized randomly. The training procedure allows the plements traditional biological sciences: “By extending the model to adjust its weights based on a series of example in- empirical foundation upon which biology is based beyond puts for which the expected output is known, using a feed- the carbon-chain life that has evolved on Earth, Artificial back error-correcting mechanism. Life can contribute to theoretical biology by locating life- The excitement for such connectionist structures which as-we-know-it within the larger picture of life-as-it-could- was growing in the 1950s received a cold shower with the be.” [18, p. 1] publication of Minksy and Papert’s forceful critique of per- Like Cybernetics in the 1960s, the field of ALife would ceptrons [20]. By showing that even simple problems are open up a whole new territory for artists. New media the- unsolvable by such linear neural networks, the book put a orist Mitchell Whitelaw remarks that ALife is an area of halt to the non-symbolic and distributed approach which experimental science which is less preoccupied by observa- had great attention in the field since the 1940s. The fund- tion and representation than it is by intervention and action. ing switched sides and for two decades, AI research turned Tracing through the interests for synthetic life in art history, towards the symbolic and heuristic approach pioneered by he hypothesizes that “a-life art” might just be the latest ad- Minksy, Papert and Simon, which would later be known as dition to “a modern creative tradition that seeks to imitate “classic AI” or Good Old Fashioned AI (GOFAI). not only the appearance of nature but its functional struc- tures” by using or appealing to technology. ALife might Systems Aesthetics then just be the true destiny of art and the realization of Jack Burnham’s vision of a “living, cyborg art form”. [31, Whereas the impact of the advent of computer science on p. 19] Western societies in the 1940s and 1950s has been thor- oughly documented, often overlooked is how it affected the Machine Leaning artistic world. In 1961, Roy Ascott’s fascination for Cy- bernetics made him envision a new conception of art as In parallel, part of the people in the AI community had be- embodied in interactive systems rather than in physical ob- come interested in questions of learning systems [17, p. jects. As a replacement for “visual art” which has become 275], paving the way to the institutionalization of a new too narrow to describe the new paradigm he attempts to de- research field within AI that would employ mathematical scribe, Ascott suggests the name “behavioural art” which models to classify and make predictions based on data or he defines as “a retroactive process of human involvement, experience rather than on logical rules.