Modeling of Culturally Affected Human Behavior
MICHAEL VAN LENT, MARK CORE, STEVE SOLOMON, MILTON ROSENBERG, RYAN MCALINDEN, AND PAUL CARPENTER
Institute for Creative Technologies
University of Southern California
Los Angeles, California
One major textbook divides artificial intelligence (AI) systems into two categories, those that are meant to think and act rationally and those that are meant to think and act like people (Russell and Norvig, 2003). Developers of systems that focus on rational behavior seek to maximize expected rewards based on the agent’s goals; they do not attempt to account for the many social or cultural factors that influence human behavior. Even systems with human-like behavior as their goal, such as ACT-R (Anderson and Lebiere, 1998) and SOAR (Laird et al., 1987), generally focus either on passing the Turing test or simulating detailed aspects of cognition and/or neural physiology. These models are becoming more like intelligent software agents and are advancing our understanding of the neural and cognitive layers of human thought. However, very few AI systems explore the sociocultural layer of human thought.
For decades, the so-called “soft sciences” (psychology, sociology, and anthropology) have focused on the behavioral phenomena that “make us human,” such as emotion, personality, and culture. Not surprisingly, however, these investigations have generated many competing, hotly debated theories but few clear answers.
A few brave AI researchers have waded into the fray by selecting individual theories and attempting to model them computationally in combination with agent-based AI methods. For example, the PsychSim system is a multi-agent-based computer program that lets social psychologists and sociologists run experiments in simulation (Pynadath and Marsella, 2005). PsychSim draws on theory of mind from the social sciences and combines it with standard AI approaches, such as decision theory and partially observable Markov decision processes, to create a system of agents that can model and reason about each other.
Similarly, the emotion-modeling and adaptation, or EMA, model (Marsella
and Gratch, 2006) combines appraisal theory (Scherer et al., 2001) with standard AI plan representations to model relationships between an agent’s goals and emotions. Finally, the cultural-cognitive architecture (Taylor et al., 2007) builds on schema theory (D’Andrade, 1992) and uses the SOAR architecture to model culturally specific behavior schemas or scripts.
Culture is one of the more complex and interesting elements of human social behavior. Even the definition of the term “culture” is hotly debated. Geert Hofstede (1994), a widely known researcher of culture, defines it as “the collective programming of the mind that separates one group of people from another.” Following in this vein, we define culture more specifically as the aspects of physical appearance, internal knowledge, and external behavior common to a cultural group. A cultural group is defined as a group of people who identify with the group through a shared trait, such as gender, race, ethnicity, religion, nationality, “regionality,” age, economic status, social class, education, or occupation.
Individuals belong to many different cultural groups simultaneously. Current theories of culture propose that, in any given situation, an individual selects a dominant cultural identity trait, which plays a primary role in influencing his or her behavior (DiMaggio, 1997). In the Culturally Affected Behavior (CAB) Project at the University of Southern California, we distinguish between culture and personality. Culture denotes the aspects of appearance, reasoning, and behavior that are common to a group; personality denotes the aspects of appearance, reasoning, and behavior that are specific to an individual and by which that individual defines his or her identity within the group.
Previous research on culture in the fields of psychology, sociology, and anthropology can be divided into two general categories. In one category, researchers (e.g., Hofstede) attempt to identify the high-level cultural parameters (such as power distance, individualism, masculinity, uncertainty avoidance, long-term orientation) that characterize a culture. This is the cultural equivalent of the Myers-Briggs personality test (Myers, 1962). Researchers in the second category focus on detailed aspects of culturally influenced behavior (such as greetings and polite or impolite gestures) (D’Andrade, 1992; DiMaggio, 1997).
Unfortunately, it is not feasible to derive low-level details of culturally influenced behavior solely from high-level cultural parameters. Hofstede’s five dimensions of culture do not provide enough information for us to derive, for example, that in Muslim cultures women should not initiate a handshake when greeting a man. However, Hofstede’s masculinity dimension (which is generally very high in Muslim cultures) could be used as an indicator that culturally important details are involved in a woman greeting a man. The masculinity dimension might also suggest that Hindu cultures, which have similar masculinity values, might share many details in this area with Muslim cultures.
Thus high-level theories can provide useful indicators and parallels that could make it less difficult to create cultural-behavior “modules” that encode the details of culturally influenced behavior. At the very least, high-level theories indicate
areas of behavior that are likely to have culturally specific aspects and/or suggest commonalities among cultures, suggesting that detailed cultural information might be reusable.
THE CULTURALLY AFFECTED BEHAVIOR PROJECT
At the CAB Project, we draw on the schema theory and theory of mind (mentioned above), plus shared symbol theory (D’Andrade, 1984), to develop a computational approach for representing, encoding, and using cultural knowledge at the individual and aggregate levels. A number of aspects of this problem are described in more detail below.
Computational. Unlike most research in sociology, anthropology, and psychology, we are attempting to develop cultural models and representations that can be integrated into an AI system that operates in an educational game. To be implementable in a computer simulation, the approach to cultural modeling must be computational. Although this approach must be fairly formal, models can still have significant uncertainties and approximations.
Approach. The goal of the CAP Project is not to develop a specific software implementation but to develop a conceptual approach that can be implemented in a variety of ways using general programming languages or AI.
Representing. One of our primary challenges is to create a representation that is not only easy to author and modify, but is also capable of supporting changes to an AI character’s cultural model without requiring that the character’s entire behavior set be re-authored.
Encoding. In addition to defining the representation of cultural knowledge, we will explore how these cultural representations can be created. One potential advantage of the modular approach to cultural representation is that it supports easier authoring of cultural models. In addition to direct authoring, encoding includes acquiring cultural knowledge from other sources, such as machine learning and data mining from databases compiled by sociologists and anthropologists.
Using. Exploring the representation and encoding aspects of cultural models requires considering how these models might be used to affect the appearance, reasoning, and behavior of AI agents. As part of the CAB Project, we are investigating a variety of ways modularized cultural models can affect agent behavior, including (1) selecting which aspect of an agent’s cultural identity should be most important in a given situation, (2) filling in details about another agent or human avatar’s cultural identity, (3) establishing opinions and attitudes toward others, (4) influencing the selection of goals and objectives, and (5) selecting culturally specific behaviors and actions for achieving those goals.
THE CAB APPROACH
Our goal is to combine established theories from the social sciences with computational methods from AI. However, we also draw inspiration from many social-science theories and focus on swappable culture “modules” that allow a virtual character in an educational computer game to look, sound, and act differently based on the currently loaded culture.
A primary hypothesis of the CAB Project is that it is possible to separate cultural knowledge from task or domain knowledge. This means an AI character’s culture can be changed without changing the character’s entire knowledge base. We are trying to modularize the cultural model into a “chunk” of knowledge that affects the AI agent’s appearance, reasoning, and behavior but is as separate from the rest of the agent’s behavior model as possible.
A central challenge is to determine the extent to which cultural knowledge, which obviously has a pervasive influence on behavior, can be modularized. Some aspects of a game character, such as appearance, can be changed very easily by switching the three-dimentional character model and texture or “skin” on that model. Slightly more challenging is modifying a character’s animations (how the character moves) and voice model (how the character sounds, including accent).
The greatest challenge, however, is modifying how a character thinks and reacts in interpersonal situations. We model a set of interlinked “sociocultural norms,” such as “is-observant-of-Islam,” “avoids-alcohol,” and “is-not-corrupt,” each of which has (1) a culturally dependent weight based on the importance of that norm to the target culture and (2) a situation-dependent degree based on how much the agent feels the current situation supports or challenges that norm. Many sociocultural norms represent “shared symbols” in that they reflect common attitudes toward specific objects, gestures, and concepts in the character’s environment.
In any given situation, the weighted sum of degrees across all sociocultural norms represents the character’s “sociocultural comfort level” with that situation. For example, Farid, the Iraqi police officer, might have a high weight for the “is-observant-of-Islam” sociocultural norm, which necessitates a high weight for the “avoids-alcohol” norm. If the human player were to offer Farid alcohol as a gift, the weight of the “avoids-alcohol” norm would decrease, and Farid would be less comfortable.
However, Fritz, the German police officer, might have a zero weight for both the “is-observant-of-Islam” norm and the “avoids-alcohol” norm. Thus offering a gift of alcohol to Fritz might seem to be a positive relationship-building action. However, if Fritz has a high weight for the “is-not-corrupt” norm, offering him any gift might be interpreted as an attempted bribe, which would challenge his “is-not-corrupt” norm and consequently decrease his sociocultural comfort level.
By weighting norms, every character, no matter what his or her culture, can include the entire set of sociocultural norms. If a norm is not relevant to a character’s specific culture, it can be “turned off” by setting the weight to zero (as with the “is-observant-of-Islam” norm for Fritz). Therefore, changing a character’s culture is simply a matter of changing the weights on the sociocultural norms rather than significantly restructuring the character’s knowledge base.
The ELECT BiLAT immersive-training application shows the advantages of the CAB cultural-behavior modules. The ELECT BiLAT application (see Figure 1) will give students an opportunity to prepare for and conduct meetings and negotiations in cross-cultural settings. At present, the cultural behaviors being developed for the ELECT project are specific to Iraqi culture. Thus the elements of Iraqi culture are not structured as a swappable module in the system but are dispersed throughout the system in both explicit and implicit ways. As a result, moving the ELECT BiLAT application to a new culture will require re-authoring
much of the system rather than just re-authoring the parts specific to culturally affected behavior.
Nevertheless, by exploring the challenges to representing and authoring cultural behavior modules, the CAB Project might make it possible to adapt systems like ELECT BiLAT to a wide variety of cultures without requiring the creation of completely new databases of behavior knowledge. We are currently working on a modified version of ELECT BiLAT that can support a meeting with either Farid, the Iraqi police officer, or Fritz, the German police officer, by modifying the weightings of the character’s sociocultural norms and changing the character’s appearance, animations, and voice model.
This example shows how cultural information might influence the behavior of a single entity. The approach works well for modeling the influence of culture at the individual level and for small groups of individuals and enables users to interact with those groups or individuals in real time. However, at the macro level, which involves the behavior of large groups and populations, the models must apply not to individual entities but to trends in behavior across interacting cultural groups. An example of culturally affected behavior at this level is the social structure of traditional tribes, which varies from culture to culture. So far, our efforts have been focused primarily on cultural behavior at the individual level, but we will also investigate macro-level cultural behaviors.
METRICS FOR EVALUATING THE SYSTEM
The first challenge to defining metrics for success in the modeling of culturally affected behavior is determining exactly what should be measured. We have attempted to evaluate four aspects of the system. First, to ensure that the system can support an immersive user experience and that it acts in a culturally appropriate way, we must evaluate the believability (also called observational fidelity) of the generated behavior. Second, to ensure stability, we evaluate the functional validity of the system (i.e., its ability to run without bugs or crashes).
Third, to ensure that new cultures can be rapidly encoded, we measure the “authorability” of the cultural knowledge modules. Finally, we evaluate the explanatory fidelity of the system (i.e., its ability to explain why the system is behaving in a given way). Metrics for believability, authorability, and explanatory fidelity will be based on acceptance tests with users. Functional validity can be measured through unit and functional testing.
The development of computational models of the social layer of human behavior is a challenging goal. Projects like CAB require teams of researchers and authorities in disciplines that have not traditionally worked together. Fortunately, we can draw on the experiences of previous generations or researchers who have
created new fields of study on the boundaries between established scientific disciplines. For example, cognitive modeling, the computational modeling of human cognition, requires expertise in cognitive psychology, computer science, and computational linguistics (the study of language using computational models). Following in the footsteps of these pioneers, computational social science has the potential to grow into a mature field of study that supports further research investigations and has important practical applications.
Anderson, J., and C. Lebiere. 1998. The Atomic Components of Thought. Mahwah, NJ: Lawrence Erlbaum Associates.
D’Andrade, R. G. 1984. Cultural meaning systems. Pp. 88-119 in Culture Theory: Essays on Mind, Self, and Emotion, R. A. Shweder and R. LeVine, eds. Cambridge, UK: Cambridge University Press.
D’Andrade, R. G. 1992. Schemas and motivation. Pp. 23-44 in Human Motives and Cultural Models, R. G. D’Andrade and C. Strauss, eds. Cambridge, UK: Cambridge University Press.
DiMaggio, P. 1997. Culture and cognition. Annual Review of Sociology 23:263-287.
Hofstede, G. 1994. Cultures and Organizations: Software of the Mind: Intercultural Cooperation and Its Importance for Survival. London: HarperCollins.
Laird, J. E., A. Newell, and P. S. Rosenbloom. 1987. SOAR: An architecture for general intelligence. Artificial Intelligence 33(1):1-64.
Marsella, S., and J. Gratch. 2006. EMA: A computational model of appraisal dynamics. Pp. 601-606 in Cybernetics and Systems 2006, R. Trappl, ed. Vienna: Austrian Society for Cybernetic Studies.
Myers, J. B. 1962. The Myers-Briggs Type Indicator Manual. Princeton, NJ: Educational Testing Service.
Pynadath, D. V., and S. C. Marsella. 2005. PsychSim: Modeling theory of mind with decision-theoretic agents. Proceedings of the 19th International Joint Conference of Artificial Intelligence. Edinburgh, Scotland, July 30-August 5. Online. Available at <http://www.ijcai.org/proceedings05.php>.
Russell, S., and P. Norvig. 2003. Artificial Intelligence: A Modern Approach, 2nd ed. New York: Prentice Hall.
Scherer, K. R., A. Schorr, and T. Johnstone, eds. 2001. Appraisal Processes in Emotion: Theory, Methods, Research. London: Oxford University Press.
Taylor, G., M. Quist, S. Furtwangler, and K. Knudsen. 2007. Toward a hybrid cultural cognitive architecture. Paper presented at the Workshop on Cognition and Culture, 29th Annual Cognitive Science Society, Nashville, Tennessee, August 1-4. CD-ROM. Available from Lawrence Erlbaum Associates.