Predictive Models of Cultural Transmission M. Afzal Upal Cognitive Science Occidental College 1600 Campus Road Los Angeles, CA, 90041 Email: [email protected] Phone: 323-259-2877

Introduction Intuitively, most people seem to understand the term 'culture' as it is used in everyday conversation1; however, it remains a notoriously difficult concept to pin down precisely. A 1952 review identified 164 definitions of culture (Kroeber & Kluckhohn 1952) and the situation has not improved since. Modern cultural scientists often resort to metaphors such as an onion or an iceberg to define culture. The idea is that culture is a hierarchy consisting of multiple layers, many of which are hidden from view. For instance, Hofstede’s cultural onion (Figure 1) consists of publicly observable symbols−gestures, pictures, words/jargon, hairstyles, and flags−as the outermost layer. Heroes−idealized people, dead or alive, seen as possessing highly prized characteristics−form the next layer. Rituals−group activities seen as essential by the group but superfluous to the achievement of the actual goal, carried out for their own sake−form the third layer. The core of a culture consists of shared beliefs about how things should be.

Figure 1: Hofstede’s Cultural Onion

1

With more than 444 Million estimated hits (obtained on 12/20/2007), culture remains a widely used term in popular culture.

Each of the layers can be further deconstructed into multiple sublayers. For instance, the privately-held widely-shared beliefs of a cultural group can be further divided into beliefs about the social world, beliefs about the physical world, and beliefs about other groups, etc. Another source of complexity is the fact that aspects of at any level and sublevel are related to aspects of other levels and sublevels. Elaborating this view, Bloch (2000), argues that “cultures form consistent wholes … every element−wherever it came from−was moulded to fit in with the others because of a psychological need for integration which led to an organically patterned ‘world view’” (p. 197). Despite the complexity, understanding culture has been important for several disciplines including anthropology, sociology, and social and cross-cultural psychology. The work in these fields has contributed to our understanding of certain aspects of culture, for instance, we have several quantitative measures of cultural differences among a variety of nations (Hofstede 1994). However, this work has been criticized for failing to develop computationally predictive models of culture that would allow us to explain the macro-level cultural patterns in terms of individual level cognitive tendencies and make predictions about the future direction of a society (Gilbert & Conte 1995; Laland & Odling-Smee 2000). The challenge then is to design models that can not only account for multiple layers of culture and the rich connections between these layers, without abstracting away the complexity, but are also computationally predictive at the same time. A complete theory of culture may also be able to satisfactorily explain how cultural layers come to be formed. Historically, we know that cultural patterns seem to come like waves on a sandy beach with the new wave rearranging the lines made the previous wave. For instance, last few centuries of Western Europian art history is a story of dynamism with one trend of cultural innovaton following another. Any two waves that are temporally contiguous in history appear to have a paradoxical relationship with each other. The new trend is both defined in opposition to the old one and as a continuation and improvement of the old trend. Visual arts are certainly not the only aspect of culture to exhibit this pattern. Other cultural trends including religious doctrines, popular cultural trends, and patterns of political thought also appears to evolve similarly. Thus Lutheranism builds on Catholicism while it also reforms it. Postmodernist art builds on Modernist art while it also redefines it. Explaining these pattern of stability and change in the evolution of cultural trends is a question of central importance for social sciences. Several critics of traditional cultural theory have offered alternatives to the standard verbal and/or mathematical modeling approaches. The alternatives include: memetics and agent-based social simulation. Next, I will critically examine these alternatives and suggest a new promising approach based on a multi-agent architecture specifically designed to lead to a computational model of culture.

Memetics Memetics is the study of culture inspired by Dawkins (1989) who coined the term meme to refer to a discrete unit of cultural information that is transmitted from one mind to another analogously to the way in which a gene propagates from one organism to another as a unit of genetic information. Dawkins argued that different aspects of culture, such as different tunes, catch-phrases, religious beliefs, and clothing fashions, compete to occupy mental space, similar to the competition among genes to be included in the DNA. Thus, only those ideas that are best fit for a mind are remembered and communicated to others, becoming widely-shared cultural beleifs. While the general idea has been well received, translating it into a viable research program has run into several difficulties. The first problem is finding a universally accepted way of dividing cultural information into discrete units. Cultural information seems to be too cohesive and well connected to yield to any single way of carving it up at the joints. There is also little evidence that the human mind is a replicating machine which simply makes a copy of the information it receives from others. Instead, when cultural information, such as a catch-phrase or a folk tale, spreads from one person to another, it seems to go through complex series of filters before being reproduced. People have to integrate the new information they receive through their senses into their existing world model. The comprehension process involves a complex two-way interaction between the newly recieved information and the knowledge that an individual possesses prior to learning. The newly obtained information may result in revision of some of the previously held beliefs. Finally, an individual may decide to communicate this information to others if he/she believes that taking the communication actions serves the speaker’s goals. Thus, sometimes there may be a causal relationship between an individual receiving a message A and then uttering a message B with A and B having some syntactic and/or semantic similarities with each other but that is not universally true. Not all messages that are received are equally likely to cause transmission of future messages. Thus informational messages are are transmitted with too low fidelity to perform a gene-like role in transmission of cultural information (Sperber 2000). This makes it hard, if not impossible, to use the abstractions employed in genetic evolutionary theory or in epidemiology to devise closed form mathematical models of cultural information trasmission of any predictive value. In fact, understanding and modeling the comprehension, belief revision, and communication biases that people have may be useful to figuring out the kind of social patterns that are likely to arise at the societal level. One of the biases that people have is the bias to pay more attention to expectation violating objects and events (Schank 1979). Holders of this bias were evolutionarily favored because they were better able to identify gaps in their existing world model and take advantage of the learning opportunities offered to them by novel events and objects around them and build more predictive models of (Upal 2005a; Upal et al. 2007). A number of recent studies have shown that people do in fact better remember and recall counterintuitive ideas but that the relationship between the amount of counterintuitiveness and recall is not linear; that objects and events that are too counterintuitive are actually recalled less well (Barrett & Nyhoff 2001; Boyer &

Ramble 2001; Gonce et al. 2006). Thus the objects and events that are minimally counterintuitive i.e., they only violate expectations about one feature (such as a talking tree) are best recalled when compared with intuitive objects that do not violate any expectations(e.g., a green tree) or maximally counterintuitive concepts that violate multiple expectations (e.g., a blinking talking tree). Anthropologists (Boyer 1994; Sperber 1996) have argued that this bias results in most of the widespread religious concepts being minimally counterintuitive. Previusly, we have argued that context in which concepts are embedded plays a critical role in the memorabilty of a concept i.e., minimally counterintuitive ideas are only more memorable when the context in which they are embedded makes them expectation violating concepts (Upal 2005a; Upal et al. 2007). Thus concepts that are counterintuitive in one context may be intuitive in another context. To get attention in the new context then concepts have to be even more counterintuitive. Hence the concepts which may have been perceived as maximally counterintuitive in the original context come to be minimally counterintuitive in the new context and exploit trasmission advantages of greater memorability. This “snowballing of counterintuitiveness effect” may help explain why some maximally counterintuitive concepts such as God are found in widespread religions and how interlinked layers of beliefs come to be (Upal 2008). My main point here, however, is that in order to have a preditive theory of cultural transmission, we need to take into account people’s memory biases. Memetics, and epidemiological models of information transmission (Watts 2002) as currently formulated to abstract away these details by appealing to mathematical evolutionary models or mathematical epidemiological models are not likely to lead to predictive models cultural information transmission.

Agent-based Social Simulation (ABSS) The key idea behind agent-based social simulation (ABSS) is to design simple bottomup computer models of individuals using software modules (called agents) and allow the agents to interact with each other through a few simple interaction rules. If any social patterns emerge then it is easy to identify individual cognitive tendencies and social interactions that cause them. This allows the ABS researchers to tease apart the micro-macro causal links by carefully making one local change at a time and by analyzing its impact on the emergent social patterns. For instance, Thomas Schelling, one of the early pioneers of the ABS approach, designed 1500 agents that lived on a 500 x 500 board (Schelling 1971). The agent’s cognitive structure consisted of one simple inference rule, namely, if the proportion of your different colored neighbors is above a tolerance threshold then move to a different cell, otherwise stay at your current location. He showed that even populations entirely consisting of agents with high tolerance end up living in segregated neighborhoods. Since Schelling’s pioneering work, the ABS systems have been used to discover possible explanations of a number of social patterns. Thus we now know the local interaction patterns that can give rise to the emergence of complex patterns of social networks. If individuals prefer to establish

connections with well connected individuals then a society is likely to have scale free network structure with a few people having a large number of social connections while a vast majority have a small number of friends (Barabasi 2002). As successful as the ABS strategy has been, it has not been able to explain the emergence of complex layers of cultural patterns that characterize human societies. To understand why it is so difficult to simulate such patterns, we need to better understand the key notion of emergence better. Emergence is not magic−even though it is treated as such by some in the ABS community. Social patterns that are seen after running an agent-based simulation are a direct consequence of the internal cognitive structure of the agent’s cognitive decision-making rules and agent-interaction rules even when we cannot foresee those consequences. This means that agent-structures and their interaction rules have to have certain properties to lead to the emergence of particular social patterns. Emergent social patterns are strongly constrained by the internal agent structure and agent interaction rules. For instance, if agent memory capacity is one-bit (Bainbridge 1995; Doran 1998; Epstein 2001) then society of such agents can never have multiple beliefs. In order to have societies with complex shared beliefs, individual agents need to be able to represent such beliefs and be able to acquire and modify them. The problem is that normative knowledge acquisition and belief revision are computationally intractable and simulating even a single agent that can perform these tasks in real time is not possible, hence designing cognitively-rich multiagent simulations that can be run efficiently is one of the greatest challenges facing those interested in creating simulations of layered cultural patterns. I believe that one way to address this challenge is to house the cognitively rich agents in synthetic “toy-domains” that are just complex enough to exercise the enhanced knowledge representation and reasoning capabilities of cognitively-rich agents but not too complex to make the simulation intractable. I will illustrate this approach with the help of a synthetic domain called Multiagent Wumpus World. Before, describing this domain, however, I will talk about a cognitively-rich agent-based social simulation architecture called the CCI-Architecture that I have designed to study the transmission of cultural information.

Communicating, Comprehending, & Integrating (CCI) Agents The CCI agents attempt to comprehend the information they perceive through their sensors, integrate it with their prior knowledge and take the action they perceive as best in a given situation. The possible actions an agent can undertake include comprehension actions, speech actions, and movement actions. The CCI agents are goal directed agents that plan sequences of actions to achieve their goals. Agents attempt to build accurate models of their environment by acquiring information about cause-effect relationships among various environmental stimuli. At each instant, agents sense their environment and decide what action to take. The CCI agents are comprehension driven. They attempt to explain their observations using their existing knowledge and their causal reasoning engine. On observing an effect OE, an agent searches for a cause C that could have produced that effect. If

multiple causes are available then the agent may have to reason to eliminate some of the possible causes to select the most likely cause for the current observations. The assumed cause AC allows the agent to make some further predictions about the unobserved effects of the assumed cause. The assumed effects (AEs) deduced from ACs are added to the agent’s world model which helps the agent form expectations about aspects of the world that the agent has not observed yet. Agent may also be able to observe causes. The observed causes (OCs) allow the agent to predict the effects (PEs) of those causes. Agents also sense actions performed by other agents that are in the vicinity of the observing agent and attempt to comprehend those actions. Other agents are assumed to be intentional agents and hence causes of their actions are those agent’s intentions. The CCI agents ignore the information received from others if they cannot find any justification for it. Inferring these intentions allows the observing agent to make predictions about the future behavior of the agent. An agent A may decide to send a message M to an agent B that happens to be within listening distance if it believes that sending B the message M will result in changing B’s mental state to cause it to perform an action C which can help A achieve some of its goals. At every instant, agents consult their knowledge-base to form expectations about the future. If these expectations are violated, they attempt to explain the reasons for these violations and if they can find those explanations, they revise their world model. We have embedded the CCI agents into the Multiagent Wumpus World (MWW) domain shown in Figure 2. MWW is an extension of Russell and Norvig’s (2003) single agent Wumpus World and is inspired by the well known minesweeper game where an agent’s objective is to navigate a minefield while looking for rewards.

Multiagent Wumpus World (MWW) MWW has the same basic configuration as the single agent Wumpus World (WW). MWW is an N x N board game with a number of wumpuses and treasures that are randomly placed in various cells. Wumpuses emit stench and treasures glitter. Stench and glitter can be sensed in the horizontal and vertical neighbors of the cell containing a wumpus or a treasure. Similar to the single agent WW, once the world is created, its configuration remains unchanged i.e., the wumpuses and treasures remain where they are throughout the duration of the game. Unlike the single agent version, MWW is inhabited by a number of agents randomly placed in various cells at the start of the simulation. An agent dies if it visits a cell containing a wumpus. When that happens, a new agent is created and placed at a randomly selection location on the board.

Smell glitter Smell glitter

glitter

Smell Smell

Smell

Smell glitter

Smell glitter

glitter

glitter

glitter

glitter

glitter

Smell

glitter

glitter Smell

Smell glitter

Smell Smell

Smell glitter

glitter

Smell glitter Smell

glitter

glitter

Smell glitter

glitter

glitter

Smell glitter

Smell Smell glitter

Smell

Smell Smell

Smell

glitter Smell

Smell glitter

glitter

Smell

Smell

glitter Smell

Smell

glitter

Smell Smell glitter

glitter

Figure 2: A 10 x 10 version of the Multiagent Wumpus World (MWW) domain. This version has 10 agents, 10 Wumpuses, and 10 Treasures.

The MWW agents have a causal model of their environment. They know that stench is caused by the presence of a wumpus in a neighboring cell while glitter is caused by the presence of treasure in a neighboring cell. Agents sense their environment and explain each stimulus they observe. While causes (such as wumpuses and treasures) explain themselves, effects (such as stench and glitter) do not. The occurrence of effects can only be explained by the occurrence of causes that could have produced the observed effects e.g., glitter can be explained by the presence of a treasure in a neighboring cell while stench can be explained by the presence of a wumpus in a neighboring cell. An observed effect, however, could have been caused by many unobserved causes e.g., the stench in cell (2, 2) observed in Figure 3 could be explained by the presence of a wumpus in any of the four cells: • Cell (1, 2), 1,3 2,3 3,3 • Cell (3, 2), • Cell (2, 1), or smell • Cell (2, 3) 3,2 1,2 2,2 1,1

2,1

3,1

Figure 3: A part of the MWW.

An agent may have reasons to eliminate some of these explanations or to prefer some of them over the others. The MWW agents use their existing knowledge to select the best explanation. Agent’s knowledge base contains both the game rules as well as their world model. A world model contains agent’s observations and past explanations. The observations record information (stench, glitter, treasure, wumpus, or nothing) the agent observed in each cell visited in the past. The MWW agents use their past observations and game knowledge to eliminate some possible explanations e.g., if an agent sensing stench in cell (2, 2) has visited the cell (1, 3) in the past and did not find sense any glitter there, then it can eliminate “wumpus at (2, 3)” as a possible explanation because if there were a wumpus at (2, 3) there would be stench in cell (1, 3). Lack of stench at (1, 3) means that there cannot be a wumpus at (2, 3). Agents use their knowledge base to form expectations about the cells that they have not visited e.g., if the agent adopts the explanation that there is a wumpus in cell (2, 1) then it can form the expectation that there will be stench in cells (1, 1) and (3, 1). In each simulation round, an agent has to decide whether to take an action or to stay motionless. Possible actions include: • • •

the action to move to the vertically or horizontally adjacent neighboring cell the action to send a message to another agent present in the same cell as the agent, and the action to process a message that the agent has received from another agent.

The MWW agents are goal directed agents that aim to visit all treasure cells on the board while avoiding wumpuses. Agents create a plan to visit all treasure cells they know about. The plan must not include any cells that contain wumpuses in them. If an agent lacks confidence in the knowledge that it currently has about a critical cell then that agent may decide to ask another agent in its vicinity for information about the cell. When an agent detects another agent in its vicinity, it ranks all the cells by how confident it is of its knowledge about a cell. It has the highest confidence in the cells that it has already visited. Next are the cells whose neighbors the agent has visited and so on. Agents also rank cells by how critical it is to find out information about that cell. The order in which the cells are to be visited determines the criticality e.g., if a cell is the next to be visited then finding information about that cell is assigned the highest priority while a cell that is not planned to be visited for another 10 rounds gets low priority. The agents then use an information seeking function that takes the two rankings (confidence and criticality) as inputs and decides what cell (if any) to seek information about. Once the first agent has sent the request for information, the second agent may also request information about a cell from the first agent in turn. A negotiation between the two agents ensues and communication takes place only if the both agents find the communication beneficial. This way information about the presence or absence of treasure, glitter, wumpus, or stench can be transmitted throughout the population and after some time t, the shared beliefs among agents may come to have a certain pattern. I believe that designing progressively richer versions of CCI and MWW and studying

the impact of each local change to see how changes in agent’s internal cognitive structure cause changes in the patterns of shared beliefs is the most effective approach to developing predictive models of cultural transmission. My students and I have conducted a number of such experiments with various versions of CCI & MWW. Upal (2006) reported that the version of a 10 × 10 MWW with 10agents was most challenging for CCI agents when it contained 10 randomly distributed wumpuses and treasures compared with MWWs containing 5 or 20 wumpuses and treasures. This is the version we used in the subsequent experiments. We found that even without any communication, false beliefs generated in such a society have a particular structure to them; they are more likely to be about objects and events whose presence is harder to confirm or disconfirm. Upal & Sama (2007) reported that communication does not eliminate or even decrease the prevalence of such false beliefs. There is some evidence to suggest that in human societies, people are also more likely to have false beliefs about unconfirmable entities and events. Bainbridge and Stark (1987) made confirmability the core of their theory of religion to argue that religious beliefs are unconfirmable algorithms to achieve rewards that are highly desired by people yet cannot be obtained. Similarly, there is some evidence to suggest that many false ethnic stereotypes people have are about things that are harder to confirm or disconfirm such as the sexual practices of the neighboring tribes (Smith 2006).

Conclusions After reviewing existing approaches to cultural modeling, I argue that a new approach based on building cognitively rich agent-based models is needed if we are to have any hope of building predictive models of cultural transmission. I describe architecture of one such multiagent society in detail and describe the work we have done to this point to study the formation of cultural patterns in human societies. While agents in our model have vastly more complex knowledge representation and reasoning capabilities than any previous agent-based social simulation models and they are able to have beliefs that are linked with each other in various interesting ways, their representation and reasoning capabilities are still too limited to result in complex belief patterns such as those that characterize any religious movement (Upal 2005b). To this end, we are currently working to enhance the capabilities of our model. We believe that this approach provides the best hope for the development of predictive models of cultural transmission.

Bibliography 1. Bainbridge, W. (1995) Neural Network Models of Religious Belief. Sociological Perspectives, 38, 483-495. 2. Bainbridge, W., & Stark, R.: A Theory of Religion. New York: Lang (1987). 3. Barabasi, A. L. (2003) Linked: How everything is connected to everything else and what it means for business, science, and everyday life, Basic Books. 4. Barrett, J. L. & Nyhof, M. (2001). Spreading Non-natural concepts: the role of intuitive conceptual structures in memory and transmission of cultural materials. Journal of Cognition and Culture, 1, 69-100.

5. Bloch, M. (2000) A well-disposed social anthropologist's problems with memes in Darwinizing Culture: The Status of Memetics As A Science, pp. 189-203, Oxford, UK: Oxford University Press. 6. Boyer, P. (1994). The Naturalness of Religious Ideas: A Cognitive Theory of Religion, Berkeley, CA: University of California Press. 7. Boyer, P. & Ramble, C. (2001). Cognitive templates for religious concepts. Cognitive Science, 25, 535-564. 8. Dawkins, R. (1989) The Selfish Gene, Oxford, UK: Oxford University Press. 9. Doran, J. (1998). Simulating collective misbelieve. Journal of Artificial Societies and Social Simulation, 1(1). 10. Epstein, J. (2001) Learning to be thoughtless: Social norms and individual computation. Computational Economics, 18(1), 9-24. 11. Gilbert, N. & Conte, R. (1995) Artificial Societies: The Computer Simulation of Social Life, London, UK: UCL Press. 12. Gonce, L. Upal, M., Slone, J. Tweney, R. (2006) Role of Context in the Recall of Counterintuitive Concepts, Journal of Cognition and Culture, 6 (3-4), 521547. 13. Hofstede, G. (1994) Cultures and Organizations, New York, NY: McGrawHill. 14. Kroeber, A. L. & Kluckhohn, C. (1952) Culture: A Critical Review of Concepts and Definitions. Cambridge, MA: Peabody Museum.

15. Laland, & Odling-Smee (2000) The Evolution of the Meme in Darwinizing Culture: The Status of Memetics As A Science, pp. 122-141, Oxford, UK: Oxford University Press. 16. Russell, S., & Norvig, P. (2003) Artificial Intelligence: A Modern Approach, 2nd ed. Englewood Cliffs, NJ: Prentice Hall. 17. Schank, R. (1979) Interestingness: Controlling Inferences, Artificial Intelligence, 12: 273–297. 18. Schelling, T. (1977) Dynamic models of segregation. Journal of Mathematical Sociology, 1, 143-186 (1977). 19. Smith, L.: Sects and Death in the Middle East. The Weekly Standard, (2006). 20. Sperber, D. 1996. Explaining Culture: A Naturalistic Approach, Malden, MA: Blackwell Publishers. 21. Sperber, D. (2000) An Objection to the Memetic Approach to Culture, in Darwinizing Culture: The Status of Memetics As A Science, pp. 122-141, Oxford, UK: Oxford University Press. 22. Upal, M. (2005a) Role of Context in Memorability of Intuitive and Counterintuitive Concepts, in Proceedings of the 27th Annual Meeting of the Cognitive Science Society, pages 2224-2229, Mahwah, NJ: Lawrence Earlbaum. 23. Upal, M. (2005b) Towards a cognitive science of new religious movements, Cognition and Culture, 5(2), 214-239. 24. Upal, M. (2007) The structure of false social beliefs, in Proceedings of the First IEEE International Symposium on Artificial Life, 282-286, Piscataway, NJ: IEEE Press.

25. Upal, M. (2008) The Layers of Culture, forthcoming. 26. Upal, M., Gonce, R., Tweney, R., & Slone, J. (2007) Contextualizing Counterintuitiveness: How context affects comprehension and memorability of counterintuitive concepts, Cognitive Science, 31, 1-25. 27. Upal, M.A., & Sama, R. (2007): Effect of Communication on the Distribution of False Social Beliefs, in Proceedings of the International Conference on Cognitive Modeling. 28. Watts, D (2002) A simple model of global cascades on random networks. in Proceedings of the National Academy of Sciences, 5766-5771.