Keywords

1 Introduction

Artificial Intelligent researchers aim, through the improvement of specific models and techniques, to achieve the best solutions for specific problems such as machine learning, computer vision and computer creativity. The discussion we intend to bring on with this paper is about the possible applications of AI that underlie the development of artificial agents able to demonstrate creative behavior in an artistic environment, based on the concepts of emergence and autopoiesis. The first has roots on Latin “emergere” that means ‘bring to light’ and is defined by Cariani [1] as the emergence of new entities that in one sense or another, could not have been predicted based on what preceded them, while autopoiesis (from the Greek “auto” which means “itself” and “poiesis” which means “creation”) describes the autonomous systems, able to self-reproduce and self-regulate, while iterate with the environment. This environment iteration might unroll, only in an indirect way, changes on the autopoietic system’s internal processes and structures [2] that might lead to a deterministic-emergent transition.

According to Wilson [3], the development of algorithms and heuristics that allow computers to perform complex and sophisticated analysis or demonstrate complex behavior, as create artworks, represents some of the greatest challenges of modern scientific research. This challenge derives not only from the development of new technologies able to support the computational requirements of such algorithms, but also the need to understand the phenomenon of intelligence and creativity through new perspectives and approaches, raising new questions on this philosophical issue.

We do not intend to engage in such a subjective discussion about intelligence or creativity but rather present a framework to help artificial agent designing focused on achieving computer creativity artwork. In this approach, the system’s intelligent behavior, as described by Smith [4], requires knowledge representation and some kind of machine learning or a certain degree of autonomous adaptation to the environment. Therefore, we might ask how to design Smith’s knowledge representation in an autopoietic system and how emergence can be seen as creativity.

On the other hand, we are deeply interested in investigating if computer creativity can be simulated using a series of algorithms, mostly from AI techniques such as Artificial Neural Networks (ANN), Genetic Algorithms (GA) and Multi-Agent Systems (MAS). One possible approach might be the identification of creative behavior in intelligent systems which main goal was not necessarily creativity simulation. The identification of such creative systems might be attained through the detection of emergent results, since (as will be discussed in this paper) one of the main characteristics of creativity, as in emergence, is the occurrence of new information, forms or expressions that didn’t exist before or wasn’t programmed or expected to arise. These intelligent systems which demonstrate emergent characteristics might as well be identified as autopoietic since usually the might be able to self-regulate, like Artificial Neural Networks.

Many researches might unroll in the junction of these mentioned areas, mainly involving the concepts of creativity, consciousness, emergence and autopoiesis aiming to create or simulate through AI techniques some capabilities that are inherent to human existence. These simulations might unroll through the development of artistic-intelligent systems that are able to express themselves autonomously in artistic terms such as music or visual arts. Therefore, some artworks that approach those above-mentioned concepts will be presented in order to illustrate the presented framework.

Finally, this paper presents a brief definition of what is an artistic-intelligent system, defined here as an Artelligent System, as well as defines a list of principles that should help the development of such systems, based previous works [5,6,7].

2 Emergence and Creativity

One can also define emergence as the appearance of macro patterns due to micro processes, so we can find in nature several examples of emergence. According to Cariani [1], the main emergent events of the universe includes particles, atoms and molecules creation, in a microscale, and stars, galaxies and black holes formations in a macroscale. One may even question if the laws of physics and even time itself are emergent aspects from the evolution of the universe.

However, emergence is something broader than the mere appearance of new structures and new patterns. It also includes fundamentally new organizations of matter and information processes along with a new world cognitive point of view. In a natural context, it is clear that the emergent transitions may involve one or more of these fundamentally new formations but it does not ordinarily apply to computer models given the different context and environment in which relationships are built: cyberspace. In a binary context, the establishment of new connections and the creation of new entities demand a new approach on the subject because one might question if the emergent transitions are possible in a virtual environment, which is a deterministic system.

Kujawski [8] affirms that it is possible for something new, unpredictable, emerge from a Turing machine once we understand the difference between rules and laws. The first is a set of well-defined formal procedures while the latter represents universal conditions. There are algorithms or a set of rules behind any emergent phenomenon, regardless their nature. A good example of emergence in a simple rules system is the Game of Life, created by John Conway in the 1950s and described in [9].

In general, emergence designates a behavior that has not been explicitly programmed in a system or agent. Pfeifer and Bongard [10] point out three kinds of emergence: (i) a global phenomenon arising from a collective behavior, (ii) individual behavior as the result of an interaction between the agent and the environment and (iii) emergence behavioral from a time scale to another.

The ant-trail formation is an example of the first emergence kind. The ants, themselves, are unaware of the fact that they are forming a trail that will determine the shortest path to food. So, when observing a population (even if it’s artificial) we might focus on the dynamic emergent characteristics of this population.

A nice example of the individual behavior as the result of an interaction with the environment is the artistic installation named La Funambule Virtuelle [11], from Marie-Hélene Tramus and Michel Bret, where a virtual acrobat evolves to keep up on a tightrope, reacting to the movements of the public. The character tries to reproduce the position of the iterator while trying to stay on the rope. In this installation, through an ANN, the balancer is able to learn to remain on the rope during the user interaction. From the learned gesture, a new behavior emerges through movements that were not taught, endowing the character of what the artist calls “the ability to improvise”.

Finally, the third kind of emergence concerns time scales. They must be incorporated from three perspectives: (i) short-term, which regards current state of the mechanism, (ii) learning and development from the ontogenetic point of view and (iii) evolutionary, phylogenetic perspective. Therefore, the three time scales - short-term, ontogenetic and phylogenetic - should be considered in order to determine whether the system is able to demonstrate emergent behavior in any of these scales.

A deeper level of emergence called “epistemic emergence” involves, of course, the emergence of new perspectives intrinsically linked to the sensorial changes. The improvement or development of new sensorial organs allows an organism to evolve into another lineage, along with new world perspectives. This kind of development also occurs in our technological evolution as we build artifacts such as thermometers, clocks, telescopes, and that extend our senses or reactions as an extension of our natural biological functions.

The installation Bacterial Orchestra (2006) by Lübke and Cornéer [12] is a good example of a creative emergence not declared by artists, expressed through autonomous objects/artifacts where an emergent phenomena emerges from a collective behavior. It consists of an orchestra formed by several autonomous cells able to hear and reproduce the sounds of the environment. The sound material comes from the ambient sound where the cells are immerse, such as people talking, sound of steps or sounds that other cells reproduce. Thus, together, they behave as a more complex organism working on an weak-defined and open domain.

Each unit of this ecology is a simple system with a microphone and a speaker. The cell is initialized randomly with a set of parameters encoded in its chromosome that will determine how it will respond to sound stimuli. The simple interaction between the cells results in a kind of microphone effect that enables new sound evolutions. These evolutions are capable of generating sound patterns that were not predicted, creating a more complex sonic space over time as the sounds reverberate through each of the cells in a kind of feedback, making it difficult to limit the scope of possibilities presented by the installation.

2.1 Creativity Identification and Classification

According to previous works [5,6,7, 13] it is possible to consider emergence as a heuristic to creativity but since there are several levels of emergence it is necessary to use some grade scale in order to classify from a lower-level emergence to a higher-level emergence. From psychology point of view creativity might not be defined as something that can be seen in a binary way - creative or not creative - but rather presents itself as a result of several variables and could be categorized according to their “creativeness”.

As for Artificial Intelligence, there is also no consensus as to the definition of creativity to this day (and there might never be). However, there are some definitions that converge in the sense that there is in the creative process some emerging function of re-signifying or creating something new, implying in the reconstruction of the past or reinterpretation of the present [14,15,16].

Although it is clear that there are several levels of creativity, there is a categorization in two main creativity levels: little-C and Big-C [17]. This type of categorization allows us to evaluate the different algorithms according to the potential generation of emergent and creative results. When comparing theoretical conceptions this quantitative distinction between little and Big-C is necessary. Big-C refers to unambiguous examples of creative expression, such as the Miles Davis Jazz or Picasso’s paintings. In contrast, little-C focuses on daily everyday creativity, for example when a person develops a new way to cook a recipe when there is a necessary ingredient missing and later receives compliments for the new recipe.

Like most dichotomies, however, this approach lacks a certain degree of softness for cases at intermediate levels. Paradoxically this approach may seem overly inclusive in some cases and non-inclusive to others. For example, if we compare three people: (i) a non-eminent artist who works professionally with the teaching and selling of watercolors, (ii) an amateur watercolor painter who uses his free time to paint, and (iii) a high school student who likes to paint sporadically. Each case qualitatively exhibits different levels of creativity, although none of them can be characterized as Big-C (if we compare with Cézanne or Kandinsky, for example).

In this sense, should the three cases mentioned above be included in the same category? By grouping we can obscure potential differences between subcategories. One way to address this kind of limitation is to create more restrictive categories with more precise cuts following “clear” examples of creativity. However, in making sharper cuts there is the risk, already mentioned, of excluding potential creative manifestations of a more subjective nature.

In order to mitigate this limitation in the traditional dichotomy, we can consider two new categories: mini-C and Pro-C [18]. The mini-C category helps differentiate subjective and objective forms of creativity that would fit into the little-C category, opening space for more subjective, personal, internal, mental or emotional forms of creativity [16]. The Pro-C category helps distinguish the fuzzy area between little-C and Big-C. Pro-C makes room for “professional” creators (such as professional artists) who have not yet attained (and might never attain) eminent status, but are still far beyond little-C creators (such as hobbyists, for example).

So, using the four-C categories proposed by Kaufman and Beghetto [18] we can classify creativity from a low-level to a higher level: mini-C representing changes in our understanding that cause impact on individuals, little-C for everyday creative thoughts and actions in any aspect of our lives which impact on individuals and their influence zone, Pro-C for creative acts of experts and experienced agents within a community or domain which impacts in a community or system and Big-C to represent eminent exceptional creativity which impact on culture, society and the world. Although there are still gaps and the number of categories is not sufficient to describe all possible levels of creativity, the four categories will be used in this research to help describe the creative potential of the algorithms.

Furthermore, Rhodes [19] developed a research that aimed to identify the multifaceted creative construction as approached in this paper. According to Rhodes, there are four Ps: Person, Process, Product and Press/Place. These Ps as used to describe the main factors that are involved in novelty creation and each component describe a fundamental aspect of a creative process. Person component includes the cognitive abilities, biological and personality bias of the creative agent which operate to a Process to create ideas, which include the stages of preparation, incubation, illumination and verification. Product represents ideas expressed in form of language, object or any other creative final outcome of the creative process. The environment and its relationship and co-evolution with the creative agent is described in Press/Place.

Together these aspects of creative creation help to describe the creative context in which a given creative outcome. This is quite important when using algorithms because it can help to describe not only how the system behave in a holistic way but also understand which are the main factors, agents and components involved. In this sense, creative outcomes are a result of creative processes engaged by creative agents which are then supported by a creative environment. More recently an extended version of this framework has been presented adding two more Ps: Persuasion and Potential [16].

3 Autopoiesis and Knowledge Representation

The concept of autopoiesis, as the organization of the living, originated in the work of Chilean biologists Maturana and Varela in the 1970s [2]. This idea was developed in the context of theoretical biology and was early associated with the artificial life simulation long before the term “artificial life” have been introduced in the late 1980s in [20].

Today the concept of autopoiesis continues to have a significant impact in the field of artificial life computing. Luisi presents a good review in [21]. Furthermore, there was also an effort to integrate the notion of autopoiesis to the field of cognitive sciences.

To be more precise, an autopoietic system is organized as a production processes network of components (synthesis and destruction) which: (i) continuously regenerate themselves in order to form a network able to reproduce components and (ii) this network constitutes the system as a distinct unit in the domain in which it exists. In addition to these two explicit criteria for autopoiesis, we can add another important point: that identity self- constitution implies on the creation of a relational domain between the system and its environment. Froese and Ziemke describe this relational domain in [22]. This emergent domain is not predetermined but possibly co-determined by the system and environment’s organization, Fig. 1. Any system that meets the criteria for autopoiesis also generates its own domain of interactions while its identity emerges.

Fig. 1.
figure 1

Graphic representation of an autopoietic system cognitive co-emergence, simplified from [22]

Full size image

A single cell organism, Fig. 2, is a perfect example of a paradigmatic autopoietic system and illustrates the circular production network that is inherent to the autopoietic self-production system. In the unicellular case, this circular relationship is expressed by the co-dependence between the limits determined by the membrane (external) and the metabolic network (internal). This metabolic network builds itself and distinguishes from the environment as a unified system. This bounded system formation is only possible due to the external system (membrane), which prevents components from dispersing in the environment. On the other hand, this external system is only constituted because there is an internal functional metabolic network. This whole system might be artificially reproduced by AI techniques such as ANN and GA.

Fig. 2.
figure 2

adapted from [21]

Single cell organism self-regulation cycle,

Full size image

The concept of self-organization can be interpreted in many different ways, but in terms of autopoietic is worthy of being presented by two aspects: (i) determining local-to-global, so that the process has its emerging identity global constituted and constrained as a result of local interactions and (ii) determining global-to-local and global identity where its ongoing contextual interaction constrain local interactions [23].

Finally, autopoietic systems are also autonomous systems since they are characterized by such a dynamic co- emergence but are specified within a specific domain. It is important for the creativity of a system that its changes and adaptations of the internal mechanisms are not performed directly by an external agent, but through an internal self-regulation mechanism.

We do not intent to present a review on the foundations of knowledge representation. Such review is widely offered by Lakemeyer and Nebel in [24]. We will assume that the agent’s intelligent behavior requires knowledge acquisition, storage and processing. To make it possible, it is essential to represent it. According to Rich and Knight [25], knowledge must be represented in such way that: (i) capture generalizations, identifying and gathering relevant properties, (ii) be understandable for people who provide it, (iii) be easily modifiable to allow error correction, reflect environmental changes, (iv) can be used in different situations even if incomplete or inaccurate, (v) help to overcome their own data volume, helping to limit the number of possibilities that should be considered.

To the machine, this symbolic pattern should be consistent enough to generate an abstraction of the domain where it is embedded. This abstraction allows it to perform operations on these patterns in order to achieve problems potential solutions. This set of symbolic patterns, in turn, may alter its collection of patterns, which consist in the agents knowledge base, through internal processes, in an autopoietic way. It means that its internal processes, self-contained in the autopoietic machine, can only change the internal organization of this set of symbolic patterns. We might say, relying on Maturana and Varela, that the autopoietic machine is a self-homeostatic system that has its own organization as a variable that remains constant. The autopoietic organization means that processes concatenated in a particular manner such that these processes produce the components of the system and specify it as a unit.

Craik [26] specified three fundamental steps for defining an agent-based knowledge: (i) the stimulus must be translated into an internal representation; (ii) cognitive processes manipulate the representation to derive new internal representations; (iii) these internal representations are translated into stimulus.

Most of the techniques found in literature represents knowledge explicitly through abstractions and use some kind of heuristic to achieve intelligent behavior. However, alternative approaches to GOFAI (Good Old Fashioned Artificial Intelligence) or classic AI, such as ANN and GA, are interesting because they bring other non-explicit knowledge representation possibilities. We should highlight that even though non-explicit knowledge is used, disregarding the need for logic, syntactic or semantic knowledge structuring, it also needs to be structured in some way. We might, therefore, consider how the agent will be able to make its own infers, alter its owns perceptions and iterate with the environment, as a circular production network.

4 Creativity Models and Computer Algorithms

In addition to the categorization of levels of creativity, Kozbelt [17] proposes a categorization of creativity theories, organized in 10 categories, which stand out for their convergence with the concepts presented in the development of this research: Developmental, Stage and Componential Processes, Evolutionary and System.

4.1 Developmental Approach

The theories related to the developmental approach are interesting to this research because they facilitate the understanding of how to plan favorable environments so the creative potential is reached. In this sense, this approach emphasizes the creative aspects of Person, Place and Potential and their results can range from mini-C to Pro-C. Although the Products do not occupy a prominent place in this approach, they play an important but tacit role. This implicit participation of the Product happens because this theory considers that there is a temporal trajectory that begins with more subjective forms of creativity (mini-C) and evolves to more tangible and mature forms of creative expression [17].

This temporal relationship is of significant importance in the context of this research because if we analyze an initial generation of random genetic algorithms, we cannot say (yet) that any creative process is expressed. However, after a few generations we can begin to see some progress regarding emergent creative processes. Another relevant aspect is the interaction between the agent and the environment where this interaction is responsible for the evolution of the creative process itself. Through this interaction the creative agent is cognitively evolved and therefore creativity could be considered as a co-evolutionary outcome of this environment-agent interaction.

4.2 Stages and Componential Processes

The theories that are fit in this approach are based on the four-stage process described by Wallas [27]. The initial stage is preparation (i) where the individual gathers information about the environment and defines a problem or objective that must be solved. Subsequently, there is the incubation process (ii) which involves a certain temporal dedication to distance oneself from the problem and dedicate itself to the cognitive process of understanding it. If the second stage is effective, then we have what Wallas considers illumination (iii). In this third stage a solution or idea presents itself to the individual (or is discovered by him). Finally, we proceed to the verification (iv) stage where the individual actually applies the solution, executes the idea and verifies the possible implications.

However, the model proposed by Wallas suggests a linearity that can hardly be verified but proposes a model of processes that if applied recursively in iterative and incremental cycles executed several times can help to refine a potential idea in order to make it increasingly adapted.

It is important to highlight that the four stages described from the human cognitive point of view find perfect equivalence in the stages found in simple reflex agents [28], Fig. 3. For an agent, the first stage (i) is the acquisition of knowledge, where it must perceive the environment through its sensors, for example. In the second stage (ii), the agent uses this representation of the world to develop some kind of reasoning. In the third stage (iii), the agent decides which is the “best” response/action. Finally, the agent performs this action (iv) through its actuators and returns to the first stage to iteratively start the process all over again.

Fig. 3.
figure 3

adapted from [28]

Model and simple reflective agent algorithm highlighted the processes of (a) preparation, (b) incubation, (c) illumination and (d) verification

Full size image

4.3 Evolutionary

According to Kozbelt et al. [17], several researches like Lumsden, Simonton and Johnson-Laird have proposed several theories of creativity based on Charles Darwin or Jean-Baptiste Lamarck biological evolution ideas. Among these theories, the Dean Simonton research [29,30,31,32,33,34] stands out as strong candidate for the most comprehensive, generally speaking, Darwinist model from a psychological point of view.

The basis of the Simonton model is, in fact, the two-stage process described by Campbell in [16], involving the “blind” generation of a set of ideas/hypotheses and the retention and selective elaboration on this set. From this point of view, ideas are randomly combined [34], below the level of consciousness. The most interesting combinations are then consciously elaborated in order to produce creative products which in turn are judged by other individuals.

Campbell’s argument for a sophisticated quantitative model of how creative productivity develops during an agent’s ontogenetic process resembles the algorithms described in evolutionary computing, such as GA, and has an important impact on understanding the eminent nature of creative processes and environments. This model assumes that there are initial differences in the creative potential between different individuals. In this sense, over time, a certain agent is able to expand his creative potential through, mainly, the exercise of the process of creation and learning from this process [32]. This theory, in a way, is consistent with described in the developmental approach.

4.4 Systems

A broader and more ambitious way of approaching creativity is to characterize it as an emergent result of a complex system that contains several subcomponents interacting with each other. Each of these subcomponents should be considered in order to propose a richer and more meaningful understanding of creative processes and results. The systemic approach proposes a more qualitative and contextualized view of creativity, almost counteracting the evolutionary quantitative view [17].

This approach emphasizes a more collaborative creativity, which is highly dependent on social conditions, rather than a vision of intrapsychic processes and individual contributions. In this sense of collaborative creativity, we can consider that systems and populations of agents with “lower” cognitive power may also demonstrate a creative characteristic that emerges from the interaction that happens inside this population. Therefore, if we consider a Multi-Agent System able to develop cognitive processes as well as act autonomously on the environment, it is also possible to model algorithms that in some way demonstrate an emergent creative behavior.

5 Developing Creative Agents

There is some effort within the AI field, especially in the cognitive AI area in order to turn the agent design principles more explicit. The discussion about these principles initially proposed by Rolf Pfeifer in the 1990s has been addressed in [10, 23, 35, 36], culminating on a thorough review by Froese and Ziemke in [22], Table 1.

Table 1. Agent design principles, apdapted from [36]
Full size table

These principles are divides into two main groups: (i) design procedure principles (P-X) and (ii) agent design principles (A-X). While the first group deals with the general philosophy linked to the chosen approach, the second deals more specifically with the methodology used for the development of autonomous agents [36]. Pfeifer and Bongard [10] present how these basic principles might be extended in order to include specific insights for each area and problem related to adaptive systems, artificial evolution and distributed systems.

The emergence design principle (P-2), as defined by Pfeifer et al. [36], is extremely relevant in this research because it demonstrates the convergence of the discussed theories towards the application of emergence as heuristics for the development of intelligent systems that demonstrate “natural” behavior. This principle is shared by many AI computational approaches in the minimal sense that the agent behavior must always emerge from the interactions with its environment.

This principle states that if we intend to develop adaptive systems, we must aim for emergence. The term emergence itself is somewhat controversial but here we use it in a pragmatic sense: something not planned or predictable. By aiming to develop an emergent agent, its cognitive structure will be the result of the history of its interactions with the environment.

To Pfeifer and Gomez [23], the relationship between behavior and emergence goes far beyond simple interactions between agent and environment. Thus, in a strict manner, the behavior is always emergent since it cannot be reduced to a simplified internal mechanism: it is always the result of the interaction system-environment. In this sense, Pfeifer et al. [36] indicate that emergence ceases to be a phenomenon with discrete characteristics (that is emergent or not emergent) and becomes as a matter of “emergence level”: the less influence the designer’s choices has on the current behavior of the agent, the higher is the emergence level. The systems developed to demonstrate an emergent behavior are usually more robust and adaptive. A system, such as genetic algorithms, that specifies the initial conditions and mechanisms for development (learning) will automatically explore the environment in order to shape its cognitive structure.

Another interesting design principle A-1, named “three constituents”, highlights the importance that any autonomous system should never be designed in isolation. Froese and Ziemke [22] point out that we must consider three components of the system that are correlated: (i) the activity field or environment, (ii) the purpose and desired behavior, and (iii) the agent itself. These three components lead us to a clear intersection with the autopoietic approach. Furthermore, Froese and Ziemke also propose that in order to better understand the intelligence phenomenon we must think the agent as a holistic system rather than study its internal components in isolation. Of course, it does not invalidate the development of the components individually, but to Froese and Ziemke, if we want to achieve a greater scientific understanding of intelligence we must investigate how the adaptive behavior emerges holistically from the dynamic brain-body-world.

As a complement to A-1, the A-2 principle is also important in this research because it denotes a clear intersection with the concept of autopoiesis. A-2 proposes that in order to better understand the phenomenon of intelligence we must search for complete agents rather than the study of the internal components of the agent in isolation. Of course, it does not invalidate the development of the components alone, but if we want to better understand intelligence we must investigate how adaptive behavior emerges from the holistic brain-body-world dynamics. Still on A-2, Pfeifer and Gomez [23] point out that agents of interest must be autonomous, self-sufficient, embodied and situated in a given context.

6 Artelligent Framework

After presenting all the main concepts and principles that guided this research, we could define an Artelligent system as an autopoietic system that through the use of artificial intelligence techniques, represents knowledge in an extensive way and considering the principles that guide the human creative process is able to demonstrate results that are acknowledged as emergent in a given environment. However it is necessary to clarify each of the component parts that were discussed in this paper since it certainly won’t group any artistic system that just involves art and artificial intelligence.

From a more objective point of view we can say that an Artelligent system should comply with the following principles, that were defined during this research and are now presented in a summarized form:

  • Use an intelligent agent or a set/system of intelligent agents, considering it’s task environment;

  • Use an AI technique, to describe an implement the previous point, that facilitates the exhibition of an emergent behavior (such as GA, ANN and MAS);

  • Represents knowledge in a extensible or emergent way (explicitly or not);

  • Apply the agent design principles, minimizing the designer role in the knowledge construction of the agent as well as maximizing the role of agent’s adaptation and learning;

  • Consider at leadt two Ps: Person, Process, Product or Press/Place;

  • Be able to generate Products that could be classified, at least, as mini-C;

  • Consider at least one of the psychological approaches to creativity: developmental, stages and componential processes, evolutionary or systemic;

  • Demonstrate an autopoietic agent behavior concerning the management of its internal knowledge and cognitive structures;

  • Be able to exhibit emergent behavior (combinatory or epistemic) or demonstrate some kind of dynamic co-emergence with the environment.

To an Artelligent system it is obvious the huge AI field influence since without those techniques it would be extremely hard to propose such a framework or a system able to represent knowledge, learn, evolve and adapt. In an Artelligent system, knowledge representation plays a fundamental role since on top of this ability cognitive and evolutive processes unroll. In this sense, the role of the designer must be to provide a platform where the agent or system might be able to create through an interactive processing of it’s own knowledge.

Among the available AI techniques, we tend to highlight in this research those which are closer to biological models such as Genetic Algorithms, Artificial Neural Networks and Multi-Agent Systems. These three approaches naturally exhibit an intrinsic emergent behavior due to their biological inspiration. In GA, even though we can determinate the main evolutionary goal through a fitness function, the evolutionary process that occurs in successive generations of individual is emergent.

This emergent behavior that can be seen in the GA is similar to what we can observe in ANN, where knowledge representation is made through the regulation of the synaptic weights for each neural connection. If we create two ANN with similar topologies, with the same number of neurons and layers, and we present them with the same stimulus in the same environment we can’t state that the knowledge will be identically represented in both ANNs.

When considering this cognitive structures, able to store and process knowledge, we should consider that knowledge representation is not determinant to the achievement of emergent results but should be seriously considered according to the chosen approach. When dealing with a MAS approach we can verify that main focus is shifted from individual cognitive behavior to collective behavior construction. In order to allow emergent behavior to rise from the interaction of a set of agents it is necessary to provide some kind of explicit or non-explicit communication protocol. In this sense, we could consider that communication and it’s protocol and form are an important part of this knowledge representation to this agents.

Multi-Agent Systems are able to demonstrate some emergent self-organization and are based in the fact that each agent is autonomous and can define it’s own goals and objectives in order to determine which actions it should do. In this kind of approach, we can develop artificial agents with a very low cognitive structure (such as ants) since the emergent characteristic is in the interaction between agents of a given population, society or ecosystem.

When thinking of an agent that will be part of an Artelligent system, we must consider it autopoietic in the sense that it must be self-sufficient and autonomous to manage its own internal mechanisms, as described earlier in this paper. We must also consider the agent design principles and consider not only the internal representation of the agent but also take into account the environment and the relationships that he will establish with it. From this point of view, the designer plays a key role in correctly choosing how to represent the environment accordingly so that the agent can interact with it but at the same time the designer must be aware that there must be enough room to cognitive construction or social self-organization of agents happen in an emergent way through an autopoietic process.

In order to illustrate those principles, we would like to briefly present Zer0Footnote 1, a gameart that invites the player to enjoy a drift in a universe ruled by geometric shapes, which is described in [7]. In this multi-agent system where each agent is visually represented by a pulsating geometric shape. Each shape has an internal clock that regulates its pulses. Every time a pulse intersects with another a sound event is generated, creating the game emergent soundtrack.

There are basically two kinds of similar agents: user-controlled and autonomous. The second is highlighted in this paper, while the first is a slightly modified version of the autonomous one in order to allow the user control its movement. The characteristics of our autonomous agents are:

  • Perception

    • Position

    • Other agents (through pulses)

    • Lifespan

  • Actions

    • Pulse

    • Move

    • Stand

  • Goals

    • Increase lifespan

    • Move

    • Interact

  • Environment

    • Infinite 2D Space

    • Multi-agent

Each agent has an internal lifespan that is initialized randomly. Since the lifespan decreases, the individuals aim to expend their lifespan though the interaction with other shapes. Every time their pulses collide, both agents increase lifespan and earn points. The larger the amount of points, more geometrical “sides” the agent has. The user agent starts with one side (a single line), than evolves side by side: triangle, rectangle, pentagon and so on.

As we can see at Fig. 4, the agent perception is based on the perception component. This component informs the agent how is the world right now, including other agents that are nearby, its actual position and lifespan (internally represented). Based on this set of information it updates his internal world representation and then the inference machine reasons which action might be suitable.

The inference machine is based rule-based agent architecture. For example, if it is not time to generate a pulse (according to its internal clock) and there are no agents nearby, move.

Fig. 4.
figure 4

Internal autonomous agent generic architecture.

Full size image

As stated before, the agents are able to establish communication through pulses and each one of those pulses is a signal. Each time they interact, they increase their lifespan. These interactions trigger sound events, thus generating the game soundtrack. Those interactions are briefly represented in Fig. 5.

Fig. 5.
figure 5

Two autonomous and one human agent represented in TROPOS [37] early requirements initial diagram.

Full size image

This environment might be described as partially observable, since it has a finite range of environment perception. Since the world’s next stage depends on other factors than the agent’s actions, it is stochastic. Also, this environment is constantly evolving while the agent is deliberating and there is no time interval. These two last characteristics impose some time constraints since the agent must answer quickly.

In this experiment the visual representation and the game soundtrack emerge from a complex environment defined with simple rules. The agents evolve along with the environment, creating some kind of self-identity, required in order to reach an autopoietic level.

7 Conclusion

In this paper we’ve tried to show evidences that may help to clarify why the autopoiesis concept can be quite interesting for artists and scientists. Computer artists, especially, may find in this concept several technological challenges that might inspire them to produce artwork. AI theorists may find fascinating and inspiring the ontology behind what was presented. The papers that deal with interactivity, autonomy and creativity can be enriched when consider all aspects of autopoiesis and emergence.

The concept of emergence offer to the art and technology fields a heuristic for creativity. If emergence can be defined as pure novelty, then understanding the processes that lead to these events, structures, functions and emerging perspectives may be relevant to the construction of artifacts that use these processes to create newness. In this sense it is possible to design and implement algorithms based on natural emergent processes in order to expand human creativity or construct artificial systems capable of demonstrate autonomous creativity.

Although it seems clear that the concept of emergency offers to the field of art and technology a heuristic for creativity, it was necessary to create a framework that encompassed a series of principles that could facilitate the arising of this emergence with a degree of controlled complexity. If emergence can be defined as the emergence of pure novelty, then understanding the processes that lead to these emergent events, structures, functions, and perspectives may be relevant to the construction of artifacts that use these processes to create novelty. In this sense, it is possible to design and implement algorithms based on natural emergent processes in order to expand human creativity or to construct artificial systems capable of demonstrating autonomous creativity.

To conclude, it would be interesting to list some possible challenges for future investigation. The theoretical understanding of intelligent behavior would be one of them since despite more than half a century of research in AI, it still lacks a thorough under- standing of the mechanisms that controls, facilitates or enables intelligent behavior. This research aims to clarify this issue by the light of autopoiesis and emergence as foundations for cognition and intelligence.