Agent-Based Modelling for Cultural Networks

229 1 Introduction We present a model based on cultural agents, that is, humanlike agents whose main framework of interactions is defined by its belonging to cultural networks. In this context, a cultural network is assumed to be made up of basic units in which an agent is connected to another agent via a cultural object. These objects are described as cultural because they contain a repository of information that may contribute to the agent’s survival or sense of identity. This chapter builds on previous work by the authors on culture and complexity (Sancho- Caparrini& Suárez, 2011; Suárez etal., 2011, 2012; Suárez etal., 2013) and cultural networks (Suárez etal., 2015; Suárez, 2020), and it attempts to design the building blocks for agentbased simulations of cultural ecosystems that will help us to understand changes that occur in cultural networks. Understanding changes that occur in cultural networks is critically important as the push for further connectivity of human life across the globe is provoking a transformation of the ways in which local and national cultural ecosystems interface. Traditionally, cultural networks presented two characteristics that are being challenged by digitization and globalization: (a) they were analogue networks in which physical contact and colocation of individuals and objects were critical for their formation and stability. But these features were not necessary for their expansion, because the combination of improvements in both travel technology and the technology to compress and codify information since the European Renaissance has made it easier for cultural networks to integrate and adapt objects from distant cultural ecosystems. And (b) they relied on cultural institutions (museums, libraries, archives, etc.), that is, socially recognized repositories of information that regulated the cultural exchanges of these networks, organizing them by serving as deposits of the memory of communities and controlling the rhythm by which novelties made it into institutional memory (Sancho- Caparrini& Suárez, 2011). As digitization becomes one of the main economic, social, and marketing drivers of the second decade of the 21st century, ever increasing on the heels of the COVID- 19 pandemic crisis, we foresee that two forces will dramatically change the nature and dynamics of traditional cultural networks. First, most traditional cultural networks are making a transition to become more and more digitally based, something that is difficult to accomplish and not strategically desirable for networks with analogue assets. Second, this digitization, even if it is partial, entails connecting (for instance, via social media) with digital platforms, which are alien to traditional 14 AGENTBASED MODELLING FOR CULTURAL NETWORKS Tagging by artificial intelligent cultural agents Fernando Sancho- Caparrini and Juan Luis Suárez DOI: 10.4324/9781003025245-17 Sancho- Caparrini and Suárez 230 cultural networks but are naturalized to some of the audiences interacting with them. These platforms are heavily dependent for many of their operations on several techniques of automation and artificial intelligence algorithms (van Dijck etal., 2018) that learn from and regulate the connections among users and between these cultural objects, ultimately affecting the dynamics of these networks. The sheer number and speed of these artificial intelligence–driven interactions outpace the ability of institutions in analogue networks to perform their regulatory roles as controllers and arbiters of both memory and novelty. The connection to these artificial intelligence–based systems results in the hegemony, or pervasiveness, of impermanent or ephemeral digital cultural objects. These ephemeral digital objects are distinguished by the fact that they leave few traces of themselves in the memory of traditional institutions (i.e. museums, archives, libraries), and within the emergence of new digital networks, ephemeral digital objects remain key elements for their maintenance. Artificial intelligence–based meta cultural objects (recommender engines, creators of navigation paths, etc.) in cultural networks short circuit and circumvent traditional repositories of memory and norms, creating a rewiring of the connections that made it possible for analogue cultural networks to form, stabilize, survive, and arbitrate the network. At the level of agents’ behaviours, we have decided to focus on and simulate the role of tagging and tags in hybrid cultural networks. Tagging is crucial to digital environments because of the relations formed among agents in these ecosystems through cultural objects that are semantic (based in tagging exercises) and dynamic (the same relation does not necessarily always yield the same tags), shaping what we describe as cultural ecosystems: semantic spaces of meaning, memory and identity that serve as the scaffolding for the creation of communities. A now wellestablished mechanism of participation in digital environments, tagging has become one of the main methods of interacting with digital objects such as texts, images, videos or sound. Whereas tagging happens at the level of the individual engaging in digital interactions, artificial intelligence–driven platforms harness the power of the masses participating in tagging and techniques for clustering of individuals as they form connections with objects and other individuals via tagging to control, foster and regulate the dynamics of the resulting networks. As a semantic system that helps understand the intentions of the taggers (the users or citizens in networks), platforms also cluster tags to both predict and direct their evolution and accurately predict their next behaviours. This ecosystem behaviour results in a multilevel digital ecosystem (users, objects, tags, traditional institutions, artificial intelligence–based super agents) that is predicated on the continuous naming (through tagging), creating and destroying of cultural moments (by interacting in the digital system through both tagging and nontagging) by virtue of permanent and ephemeral tagging of the world. The structure of the chapter is as follows: in section2 we present the definition of cultural networks that we use, highlighting some of the features that differentiate them from other networks that we can find in other social science problems. Next, we will describe the modelling fundamentals that will support the framework proposed in this work, emphasizing its two main cornerstones: the BDI paradigm for modelling human behaviour through agents and the random surfer framework (inspired by the functioning of PageRank algorithm) for modelling the decisionmaking and exploration of the network by those agents. We end section3 by presenting the concrete modelling framework that is the core of our work. In section4, we produce a first formalization of a model that, following the presented framework, builds an evolutionary mechanism of cultural networks that approximates the complex behaviour that we can find in some real cultural networks. Here, we also show some preliminary results that have been obtained with an implementation of this model. Finally, we finish this chapter with some Agentbased modeling for cultural networks 231 conclusions that can be inferred from this work as well as some lines of future research (some of which are already underway). 2 Cultural networks For the purpose of this chapter, a cultural network is understood to be a multimodal network in which at least two types of nodes represent people and objects, respectively, and in which a cultural object or phenomenon symbolically and cognitively connects people to one another. That is, a cultural relationship is created via an object or phenomenon that semantically loads the links connecting humans through joint attention (Tomasello, 2000) with information contextually relevant to harnessing the existing environmental complexity, ultimately facilitating the cultural learning of participants within the cultural network. Cultural networks provide the basic mechanisms required for the process of meaning making for individuals and communities; they constitute the infrastructure of symbols, tools and objects that keep communities together, providing a basis for the development of prosociality (Sancho- Caparrini& Suárez, 2011); they function as the interface used as a means to negotiate encounters between different cultures (Suárez, 2007); and they carry the requisite tools that enable cultural formations to adapt, transmute and evolve. Understanding both the concept and the dynamics of cultural networks is critically necessary within the broader fields of digital humanities, social computing and cultural analytics. Historically, within these fields there has been a lack of recognition that cultural networks must be understood as living, responsive organisms that are leading complex processes connecting the cultural history of communities with the mechanisms of cultural evolution (Cavalli Sforza, 2007; Richerson& Boyd, 2006; Henrich, 2016; Tomasello, 2019; Inglehart, 2018). As such, it is imperative that these fields recognize that cultural networks provide the ground for the development of new tools that will uniquely position these fields to link the human past, in historical terms, with the temporal scale of the evolutionary processes of culture. Before the transition to exclusively digital cultural networks has advanced, thanks to the adoption of artificial intelligence, we urgently need to understand how digitization and artificial intelligence affect and change the traditional roles and functions that analogue and hybrid networks have served for both communities and individuals. In this chapter, we intend to develop a model of digital cultural networks that helps us to understand how the introduction of artificial intelligence–based cultural agents affects the dynamics of these cultural networks. We will focus on the tagging of digital cultural objects as the main action of our humantype agents, while our artificial intelligence–based agents will influence subsequent iterations of the cultural network by presenting new sets of cultural objects resulting from the analysis of previous rounds of tagging and clustering by humantype agents. We will emphasize the change that the interactions through different levels of the network create when compared with the traditional cultural networks organized around canonical objects and held within centralized repositories of community memory. Importantly, we will highlight the impact that artificial intelligence surrogates and tagging have on cultural networks as these become less centralized, more ephemeral and better suited to satisfy the business models of digital platforms. Artificial intelligence surrogates and tagging also allow for the creation of constant cycles of cohesion (similar tags) and destruction (dissimilar tags or perceived absence of tags, which is another form of tagging in and of itself) of communities (both digital and analogue) through common cultural objects and the satisfaction of individual desires by targeted tagging of specific cultural objects. Sancho- Caparrini and Suárez 232 3 Modelling human interactions and cultural complexity 3.1 The BDI paradigm Although agentbased modelling provides one of the most fruitful tools for understanding complex systems (Siegfried, 2014), we still encounter unresolved difficulties in deciding how to shape the individual rules that agents must follow to obtain the purposeful design of local and global behaviour. This difficulty is found not only from an operational and functional point of view, about how to design the rules that affect the different entities that intervene in the modelling, but also from a conceptual and epistemological point of view, where the rules and their representation have an equivalence in the real world that the modeller wants to reproduce. And if this is true in the modelling of complex systems in general, it becomes especially difficult when dealing with the modelling and behaviour of systems in which the agents (or some of them) represent human beings. Among conceptual models, the most successful in modelling human beings, the BDI architecture (Bratman, 1987, 2009; Grosz& Kraus, 1996), is based on a vision of modelling individuals as rational agents who possess certain mental attitudes: beliefs, desires and intentions (hence the name of the architecture) which represent, respectively, the informative, motivational and deliberative states of the agent. These mental attitudes will allow, at least conceptually, the modelling of individuals in the system as rational agents but with what we understand to be bounded rationality. The concept of bounded rationality is necessary because it includes the idea of incomplete informative states and limited intentions, which are necessary conditions to confront in any model that aims to simulate human agents. Bounded rationality is much more closely aligned to reallife situations for human beings when they are faced with making decisions in normallife conditions. In BDI architecture: 1. Beliefs represents the individual’s knowledge about the environment and about their own internal state; 2. Desires represents the goals which the individual has decided they want to achieve; 3. Intentions are the set of plans (sequence of actions) which the individual intends to follow to achieve their goals. Conceptually, BDI architecture provides an approach called folk psychology, since it is based not on a physically real model of how reasoning works (at least, not in a verified way) but on how we popularly believe that the decisionmaking process takes place, at a high level of abstraction. Although its author, Michael Bratman, did not introduce this architecture as a practical means for the creation of multiagent models, both he and other authors realized the potential that this architecture had for the effective construction of real models. Interestingly, shortly afterward its creation, the BDI architecture would have more influence and application within the world of computing than in the world of philosophy, where it was created. In a short period of time, the BDI architecture gave rise to a new computational paradigm with an immediate application in the specification and implementation of intelligent agents within the area of artificial intelligence (AI). Later on, Cohen and Levesque (1990) distinguished BDI from purely reactive models by emphasizing the existence of commitments to intentions in order to achieve longterm goals, transforming the action of commitments into a critical part of the BDI model. Rao and Georgeff (1991) showed that we need suitable rational processes to decide the intentions to focus on in the long term under sets of specific circumstances. These subsequent studies have brought into real convergence areas of AI that, although intuitively close, could hardly be found together in Agentbased modeling for cultural networks 233 real projects, such as agent modelling and planning, decisionmaking, optimization and nonclassical logics (BDI logic) techniques. Since its introduction, BDI architecture has also provided a large number of applications in practical social models (Meyer etal., 2001; Adam& Gaudou, 2015; Helbing etal., 2000), enriching many of the available multiagent system building frameworks (commercial and free). We can find many variants and extensions to adapt it to several scenarios where specific features in the agents are needed (Singh etal., 2011; de Silva etal., 2009; Taibi, 2010). 3.2 PageRank, random surfers and recommender systems Generically, PageRank is the name given to the various versions of a web page ranking algorithm that was designed by Larry Page and Sergey Brin when they were at Stanford University (Page etal., 1999). In addition to the impact on the world of computation that this algorithm has for the problem it solves, it has great historical, technological and social importance because its creation gave rise to the invention of Google’s search engine. The algorithm provides a method to measure the importance of a web page within a system of pages, based on the quantity and quality of links pointing to it. The problem of measuring the importance of the nodes in a network according to how they are connected to each other is an old problem that was already known and addressed from different angles within the field. At the moment of PageRank’s invention there were wellknown solutions, even if they were not efficient from a practical point of view, when the size of the network grew minimally. Perhaps the first known practical application of this problem is due to the economist Wassily Leontief (Harvard University, and Nobel Prize in 1971) who used it to model the functioning of an economy represented by a network in order to give a ranking between the different components of the economic system. Along with the algebraic solutions to solve this ranking problem (Arasu etal., 2002), which include the various variants used by Google, we can find a very simple distributed solution that also provides a more extensible and inspiring solution based on the behaviour of what are called random surfers (Richardson& Domingos, 2002). Let’s assume an individual (our surfer) is surfing the network. Their way of surfing is completely random, following the existing connections: at a certain moment this surfer is in a node, then they select at random and following a uniform probability, one of the nodes that are connected to that node, and jumps to it. This is called a random walk through the network (hence the name random surfer). If we repeat this process many times and record the number of times the surfer is at each node, we can prove that the numbers at each node are proportional to the importance of the node in the network structure. Despite some small, easily solvable drawbacks of this method, it has some clear advantages, such as: • It is robust to changes in connections, because most walks will be minimally affected if the changes are local. • It is easily parallelizable, just by having several surfers moving through the network simultaneously. • It is distributed. It’s easy to have several independent processes (they don’t even need to be synchronized) working with different surfers. • It is local to the topology of the network, which means that if we want to have information about the importance of a part of the network, we can originally place the surfers in that part and see the local results they produce. Sancho- Caparrini and Suárez 234 In addition, it also has advantages for the generation of variants that can be interpreted as alternative analyses to the study of networks. For example: • It is easy to consider cases where the probabilities of visiting neighbours from a node are not uniform, so there may be a preference for certain visits (if we think of a cultural network, for example the rating or tags assigned in the links). • It is easy to introduce variants where surfers have preferences, so there may be some specializedin- edge tags or internal characteristics of the nodes they visit. • An importance ranking (like the one provided by the PageRank algorithm) can be obtained from a set of nodes with respect to a fixed one, making the surfer always start from a specific node. • The importance ranking of structures larger than single nodes can be estimated by considering the visits that can be counted to a set of nodes from any other set of nodes. • The importance ranking of edges, and not only of nodes, can also be estimated by recording the visits that they receive in random path moves. According to the latter considerations we can use a selective surfer system as a custom ranking procedure (even from a single node) that can provide similarity measures, clustering, recommendations, etc. 3.3 Aframework model of cultural networks From the previous building blocks (the BDI paradigm as a model of intention and random surfers as a model for decentralized selection) we have built an elegant model for cultural networks that allows us to analyze and measure how the interactions between the elements of the network (human beings and cultural objects) define the structure and dynamics of the network itself as a whole. Conceptually, the abstract framework model through which we model cultural networks is structured by several layers: 1. In the first layer we find the traditional cultural network itself, formed by the individuals (humanlike agents) and the cultural objects with which they interact. Formally, it is a network where the nodes are the humans and objects, and the edges are the various interactions that can occur between them. In a first approach we consider that the network is bipartite, in the sense that the only interactions/edges between nodes are those between nodes of different types (humans/cultural objects). Since we can introduce more or fewer attributes in the nodes and we can enrich the interactions between them, we are considering a more complex computational model than the usual mathematical graph that just stores information in nodes (attributes of humans/objects) and edges (properties of the interactions that occur). It is at this level that features of the BDI paradigm can be implemented to model the dynamics of individuals in their interactions with cultural objects (or with other individuals, if we want to add that layer of possible interactions). 2. In the second layer, we find the semantic relationships that can be found between the attributes (information stored in nodes and edges) of the previous layer. We can consider that this second layer is a second network in which the nodes are the attributes used in the previous layer, while the relations will be defined by how those attributes are used and interpreted in that layer. In fact, this second layer can be seen as a collection of several networks that Agentbased modeling for cultural networks 235 work as projections, where we obtain a different semantic network each time we capture a specific moment an interaction happens in the model, depending on the set of relationships that we want to highlight/analyze from the previous layer. 3. A third and last layer is given by auxiliary procedures that allow us to enrich the knowledge of the network that each individual can acquire. In our framework, this personal (and inaccurate) knowledge is given by a group of surfing agents that eventually go through the network and inform individuals about the features of the observed network. We can interpret them as personalized random surfers, whose behaviour is determined by the individual’s current characteristics and their relationship with the cultural objects with which they interact. Therefore, and to distinguish them from the individuals of the first layer, we will call them cultural surrogates (or surrogates in short). Their role is strictly informative, and they cannot directly modify the traditional cultural network in the first layer. From the point of view of the BDI model of individuals, surrogates are in charge of populating the beliefs of each individual. These surrogates (for instance, the algorithm of a social network that learns from the user’s behaviour in order to present them personalized content or recommendations in the future) differentiate artificial intelligence–based cultural networks from traditional cultural networks in that these surrogates are very efficient in their tagging of objects for individual cultural agents. Traditional cultural networks only had cultural institutions (museums, libraries, archives) as cultural surrogates, whose functions were to generalize the memories of the traditional cultural network, while serving the whole population of the network. Within this general framework is enough freedom to consider or to implement more or less complex models of each level of cultural networks in order to make different approaches to modelling them. We can also try a variety of approaches to test different theories about how information is organized and flows in these networks, or even how it is more or less pertinent to introduce complex models of the human being (enriching decisionmaking by means of more nuanced BDI models) so that certain behaviours emerge in the network. We must note that, although we can make a static analysis of the network based on how the connections at the three levels look at a given moment in time, one of the most interesting aspects of this framework is that it is based precisely on the dynamism inherent in cultural networks. This dynamism is reflected in the actions carried out by individuals as a result of the processes determined by their own BDI model and that can replicate, for example, the usual actions of “like/dislike” in a digital social network (such as Facebook, Twitter, etc.) that an individual can perform on another individual’s publication. The interaction of individuals with objects not only shapes a cultural network in any given moment, but also makes it evolve over time, producing phenomena that can be observed and measured at all three layers, such as • Layer 1: birth/death (creation/deletion) of content (new cultural objects), birth/death (connection/disconnection) of individuals, generation of clusters of individuals/objects sharing similar characteristics, etc. • Layer 2: dynamic grouping of interactions (e.g. certain behaviour in the network that precedes the creation/deletion of content), emergence of semantic connections between object properties, segmentation of the population based on the way they behave, etc. • Layer 3: classification of decision patterns in the network (comparing the surrogates of individuals), complexity of the information space that individuals attend to, evolution of the characteristics of the surrogates, etc. Sancho- Caparrini and Suárez 236 4 Implementation 4.1 Arestricted model Building a complete framework like the one described earlier is a difficult and longrange project. Therefore, what we present in this section is a first approach to a concrete model that follows the guidelines and fundamental ideas of the framework. It does not aim to offer all the functionalities in a complete way, but is rather a demonstration of a small agent model that replicates some of the characteristics and process dynamics that we find in some cultural networks according to our theoretical approach as described previously. We will rely on the scenario outlined in section2, where the interaction of human agents with cultural objects is restricted to connecting with those objects through computing affinity that generates a social semantic map, where surrogates surf the map or network looking for related objects to ensure the satisfaction of the individual in the network and, as such, their permanence in it. Thus, cultural objects are described through a set of tags. To simplify the implementation and later analysis, tags are handled among a finite and prefixed set of possibilities: T. We will denote A for the set of individuals and O for the set of cultural objects. In this model, we will assume that the individuals cannot tag the objects, but rather the tags are preassociated to the cultural objects. For each object in O, its set of tags is given as a subset of T. Meanwhile, individuals have a set of “preferences”, P, which stores the characteristics of the ideal cultural objects that satisfy them the most. Apreference of an individual will be a set gg g =∪ +− , where g+ and g− are disjoint subsets of T, g+ (resp. g −) indicates the tags that they desire to be (resp. not to be) present in the object. The tags not in g do not affect the perceived affinity between the object and individual. When a cultural object is presented to an individual, the latter measures how well the object fits their preferences, their affinity. We will use a simple set calculation to compute the affinity that considers the number of matches between desired and undesired tags, that is, if aA∈ , P is their preferences, and e are the tags of the object o, then: affinit yaomax ge ge ge ge gP cc , () =∩+∩−∩ −∩ () ∈ {} + −+− where e c is the complementary set of e (i.e. the tags not in o). The individual calculates the affinity with o as the best of the affinities with respect to the set of their preferences. This value is always in the range −  11 ,. When affinit yao, () = 1 , the object meets exactly one of the individual’s expectations, and when affinit yao, () =− 1 , the object is at the opposite end of their expectations. The goal of such a scenario is to demonstrate how clusters can be produced between individuals and objects through the affinity between the preferences of the former and the tags associated with the latter. This way, we have a snapshot of the structure of the cultural network that arises in this process of connecting by affinity and which corresponds to the level 1 of analysis mentioned earlier: clusters formed both in the set of individuals (by association with common objects) and in the set of objects (by association between communities of individuals). In this context, we can consider an initial cultural network, AOE∪ () ,, formed by agents of A, objects of O, and connections ao E, () ∈ if and only if affinit ya ot haff , () < (minimum threshold of affinity). Once this network is created, we will remove the isolated nodes (from A and O), since in the dynamics to be introduced they would not play any possible role (in this dynamics objects that are exposed to individuals are selected among those connected and evaluated by other individuals, or their mutations). Agentbased modeling for cultural networks 237 Broadly speaking, the dynamic we propose follows these four general steps (repeated cyclically): 1. Connection: For each individual, select an object to present to them so that the individual decides (through their affinity) whether to connect them. Update the satisfaction of individuals. 2. Birth: Add new individuals and objects to the world, by cloning and mutation of existing ones. 3. Oblivion: Remove objects with lowest influence that have not been recently presented to individuals. 4. Death: Remove individuals and objects that have become disconnected and individuals with satisfaction below a required threshold. Satisfaction of an individual is simulated through an accumulated function that decreases in time and increases every time the individual finds a new object of interest (high enough affinity). In the Connection step, the selection of objects to be presented to individuals is made by using the surrogates that each individual launches to decide which objects in the network are not connected to them and could be of interest. The Birth step tries to mimic the procedure that occurs in some cultural networks for the creation of new elements: some highly satisfied individuals invite new individuals (with similar preferences) to join the network by means of sharing or showing them some object of attraction; and, simultaneously, objects that present slight variations with respect to successful objects (high importance, for example using PageRank) are created in the network. Since we want to maintain network capacity under some constraints, the Oblivion step produces some objects that will be forgotten. This process can be also found in many cultural networks, where cultural objects with less importance in the network and not recently used by anyone end up being forgotten. Finally, unsatisfied individuals (represent individuals who have not found interesting information in the network) leave the network, and those elements that have become isolated are removed because it is very unlikely that they will be able to intervene in the following life cycles of the network. The BDI architecture is being used in this approach through the following association for each individual: beliefs are given by the direct knowledge they have about their egonetwork and by the information that surrogates provide them through the selection of new objects; desires are represented through individuals’ preferences (static) and the attempt to maximize their satisfaction in the network (dynamic); individuals’ intentions are to stay in the network, which is translated into actions through the use of surrogates looking for new information and the connection with objects that are related to their desires. 4.2 Experiments In our example, we start with a randomly generated network of 100 individuals and 100 objects (some of these elements might be disconnected from the first step and not appear anymore), which can work with a set of 100 tags (there are 2 100 possible different objects). The individuals have a set of 5 preferences with which they measure the affinity of the cultural objects they analyze (there are 3 100 possible different preferences, so there are around 3 500 possible different individuals).