Abstract
Cognitive social simulation is at the intersection of cognitive modeling and social simulation, two forms of computer-based, quantitative modeling and understanding. Cognitive modeling centers on producing precise computational or mathematical models of mental processes (such as human reasoning or decision making), while social simulation focuses on precise models of social processes (such as group discussion or collective decision making). By combining cognitive and social models, cognitive social simulation is poised to address issues concerning both individual and social processes. To better anticipate the implications of policies, detailed simulation enables precise analysis of possible scenarios and outcomes. Thus, cognitive social simulation will have practical applications in relation to policy making in many areas that require understanding at both the individual and the aggregate level.
Tweet
Blending cognitive and social models leads to tools for more precisely understanding policy implications at both individual and social levels.
Key Points
Cognitive modeling and social simulation together capture both individual mental processes and interpersonal social processes, for better understanding the implications of public policies.
Detailed simulation enables precise understanding of possible scenarios and outcomes, which can guide better policies.
Cognitive social simulation should be part of the curriculum of studying public policies.
Introduction
Predicting the effects of policies can benefit from detailed computational analyses. As a simple comparison, weather forecast has improved tremendously with the use of computer models for various scenarios producing outcomes (e.g., paths of a hurricane) with specified probabilities. Similarly, when policy makers consider a certain social or economic policy, they preferably would want to know the full implications of it. They would like to know the implications in terms of quantifiable and measurable outcomes, such as the amount of total increase in revenue or the total cost to tax payers. But, beyond these rather direct outcomes, they may also want to consider more indirect implications, such as how it affects different individuals’ perception, emotion, and motivation; how changed perception, emotion, and motivation lead to cultural and societal changes; and how all of these changes lead to altering quantifiable and not-so-quantifiable outcomes. Rather than relying on speculations, one would definitely want more reliable means for understanding them.
Thus, policy makers (and others) may need to look into not just social sciences but also cognitive sciences (broadly defined), to better understand such issues in relation to policies. Moreover, they may also want to look into combining social and cognitive sciences somehow, to connect analyses at different levels for the sake of a more comprehensive understanding (Sun, 2012).
Furthermore, rather than relying purely on verbal–conceptual theories regarding complex social and cognitive matters, a more exact, more quantitative approach should be more desirable. For example, given the complexity of the human mind, it has proven difficult to infer fine-grained cognitive-psychological details from behavior alone. Although experimentalists may come up with an informal (verbal–conceptual) theory to aid inquiry, full consequences of such a theory may not be obvious, its details may be underspecified, and its ambiguity and inconsistencies may be hard to discover or avoid (Sun, Coward, & Zenzen, 2005). Computational modeling, unlike verbal–conceptual theories, is precise and also expressive (capable of precisely describing many details). It is a suitable ground upon which detailed theories may be constructed and tested (Sun et al., 2005; Vernon, 2014). Sun (2001, 2006) argued for the role of computational modeling in understanding social-cognitive issues, especially computational social simulation with realistic cognitive models (i.e., cognitive social simulation), utilizing “cognitive architectures” in particular. The present article aims to explore such possibilities.
Specifically, cognitive modeling, an approach developed in cognitive sciences, centers on producing precise computational (or mathematical) models of individual mental processes (such as detailed models of human memory, reasoning, or decision making). Social simulation, as developed in social sciences, focuses on precise computational models of social processes (such as models of interaction between two individuals, group discussions and decision making, or other collective processes). Cognitive social simulation combines methodologies of both cognitive modeling and social simulation (see examples in subsequent sections). By combining cognitive and social computational models, cognitive social simulation is poised to address issues concerning both individual and social processes and their interaction. Thus, cognitive social simulation will have practical applications in relation to policy making in many areas that require understanding at both the individual and the aggregate level.
Note that the present article does not aim at providing specific policy recommendations. Rather, it aims at describing how some lines of research may lead, in the near future, to providing specific policy recommendations in many areas (e.g., through quantitatively exploring deep policy implications). If one is looking for any concrete recommendation from this article, it is that the further development of these lines of research will benefit future society and future policy making, and thus should be closely watched or adopted by policy makers.
Combining Cognitive Modeling and Social Simulation
Some basic concepts will be explained below that will show how this combination works. The notion of “agent” (i.e., autonomous entities) has occupied a major role in defining social and cognitive research. Below, I will briefly examine this notion in both the social sciences and the cognitive sciences, which points to their integration.
Computational models of agents often take the form of a “cognitive architecture” as developed in the cognitive sciences, that is, a broadly scoped, domain-generic computational model describing the essential structures and processes of cognition (psychology). They are often used for broad analysis of cognition and behavior (Anderson & Lebiere, 1998; Carley & Newell, 1994; Sun, 2002, 2016; Vernon, 2014). In particular, cognitive architectures provide a means for specifying a wide range of cognitive-psychological processes together, in tangible (computational) forms, although traditionally the focus of research in the cognitive sciences has been on specific aspects of cognition. For example, a cognitive architecture may include memory, reasoning, decision making, and other cognitive functionalities.
A cognitive architecture provides a concrete framework for more detailed modeling and simulation of cognitive-psychological phenomena, through specifying important structures and a variety of other essential aspects. It thus helps to narrow down possibilities, provide scaffolding, and embody foundational theoretical assumptions. The usefulness of cognitive architectures has been demonstrated and argued before (see, for example, Anderson & Lebiere, 1998; Sun, 2002, 2016; Vernon, 2014).
Computational cognitive modeling, especially with cognitive architectures, has become an essential area of research in the cognitive sciences. Cognitive architectures specify, often in considerable computational detail, the mechanisms and processes underlying cognition. Cognitive architectures unify various subfields of the cognitive sciences by providing unified computational accounts of specialized findings in an integrated model. Some of them have accounted for hundreds of phenomena from cognitive psychology, social psychology, personality psychology, industrial/organizational psychology, and more (e.g., Sun, 2016). Such developments, however, need to be extended to issues of multiagent social interaction.
In contrast, most models of agents in the social sciences have been simple, although there have been some promising developments recently (Conte, Andrighetto, & Campennl, 2013; Edmonds, 2014; Sun, 2006). Generally speaking, two approaches dominate the social sciences. The first approach may be termed the “deductive” approach (Axelrod, 1997; Moss, 1999), exemplified by much research in classical economics. It centers on the construction of mathematical models, usually as a set of equations. Deduction may be used to find consequences of assumptions. The second approach may be termed the “inductive” approach, exemplified by many traditional approaches to sociology. Insights are obtained by generalizations from observations; these insights are often qualitative, and phenomena are described in terms of general categories.
However, a relatively new approach involves computational modeling and simulation of social phenomena, which starts with a set of assumptions in the forms of rules, mechanisms, or processes. Simulations then lead to data that can be analyzed. Both inductive and deductive methods may be applied on simulation data: Induction can be used to find patterns in data, and deduction can be used to find consequences of assumptions (i.e., rules, mechanisms, and processes specified). Thus, simulations are useful in multiple ways (Axelrod, 1997; Moss, 1999).
This third approach centers on agent-based social simulations, that is, simulations based on autonomous individual entities. Such simulations explore the interaction among agents whereby complex patterns may emerge. Thus, they provide explanations for corresponding social phenomena (Gilbert & Conte, 1995). Agent-based social simulation often tests theoretical models in the social sciences or investigates their properties (when analytical solutions are difficult). A simulation may even serve as a theory by itself. Researchers have turned to agents for studying a wide range of issues (Conte, Hegselmann, & Terna, 1997; Gilbert & Conte, 1995; Gilbert & Doran, 1994; Moss & Davidsson, 2001).
Some work in social simulation assumes rudimentary cognition-psychology on the part of agents: Agent models have often been custom-tailored to the task at hand, often just a restricted set of highly domain-specific rules, not comparable to cognitive architectures in sophistication. Although this approach may be adequate for achieving some limited objectives, it is overall unsatisfactory: It not only limits the realism, and hence applicability, of social simulation but also precludes the possibility of tackling the theoretical question of the micro–macro link (Alexander, Giesen, Munch, & Smelser, 1987; Sawyer, 2003).
Investigation, modeling, and simulation of social phenomena need cognitive sciences, because such endeavors need a better understanding, and better models, of individual cognition-psychology; only on this basis can better models of aggregate processes be developed (Castelfranchi, 2001; Sun, 2001). Cognitive models may provide better grounding for understanding multiagent interaction, by incorporating realistic constraints, capabilities, and tendencies of individual agents in their interaction with their environments (e.g., as argued at length in Sun, 2001). Some researchers have already started to explore the cognitive basis of social, political, religious, and cultural processes (e.g., Atran & Norenzayan, 2004; Boyer & Ramble, 2001; Castelfranchi, 2001; Jager, 2017; Kim, Taber, & Lodge, 2010; Mithen, 1996; Turner, 2000). Although some cognitive details may ultimately prove to be irrelevant, this cannot be determined a priori, and thus modeling may be useful in determining which aspects of cognition can be safely abstracted away.
Although, generally speaking, computational modeling is often limited to a particular level at a time (e.g., the social, the cognitive-psychological, etc.), this need not be the case: Cross-level analysis and modeling, such as combining cognitive modeling and social simulation, could be enlightening, and might even be essential (Sun, 2012; Sun et al., 2005). These levels do interact with each other (e.g., by constraining each other) and may not be easily isolated and tackled alone. Moreover, their respective territories are often intermingled, without clear-cut boundaries. One may start with purely social descriptions but then substitute cognitive-psychological principles and details for simpler descriptions of agents. Thus, the differences and the separations among levels can be rather fluid. (Note that Sun et al., 2005, and Sun, 2012, provided detailed arguments for crossing and mixing the levels of the social, the cognitive-psychological, etc.; Sun, 2006, provided more technical discussions of integrating social simulation and cognitive modeling.)
The remainder of this article discusses examples of cognitive social simulation. Note that simulations may differ in terms of their cognitive-psychological realism. Social simulation models can be rather noncognitive, by using, for example, a simple set of rules for an individual agent (Axelrod, 1984). Social simulation models can also be much more cognitive, by using well-developed cognitive architectures (e.g., Sun & Naveh, 2004). In between, there can be models of various cognitive complexity (Carley & Newell, 1994; Goldspink, 2000; Jager, 2017). In terms of noncognitive details, one may include in a model only highly abstract social scenarios, for example, as described by game theory (Von Neumann & Morgenstern, 1944), or one may include a lot more details of the scenarios as captured in ethnographical studies (e.g., Clancey, Sierhuis, Damer, & Brodsky, 2006).
Examples of Cognitive Social Simulation
Below we look into a few examples of cognitive social simulation.
A Cognitive Simulation of Games
Some work in cognitive social simulation extends existing formal frameworks of agent interaction, taking into consideration cognitive processes more realistically. For instance, various modifications of, and extensions to, game theory (Von Neumann & Morgenstern, 1944) move in the direction of enhanced cognitive realism. Although policy makers may use game theory to find mathematically optimal strategies for various situations, humans often do not adopt optimal game theoretical strategies (Axelrod, 1984).
For instance, a cognitive social simulation (by West, Lebiere, & Bothell, 2006) found that human players did not use a fixed way of responding. Instead, they attempted to adjust their responses to exploit perceived weaknesses in their opponents’ play. The researchers argued that humans had evolved to be such a player; furthermore, they argued that the human cognitive system had evolved to support a superior ability as such a player.
These researchers (West et al., 2006) created a cognitive model of how people play games by applying the ACT-R cognitive architecture (Anderson & Lebiere, 1998), and then compared it with the behavior of actual human players. For instance, the standard game theory requires that players be able to select moves randomly in accordance with preset probabilities, but research has repeatedly shown that people are very poor at doing this, suggesting that the evolutionary success of humans is not based on this ability. People try to detect the opponent’s sequential patterns of contingent choices (such as tit-for-tat) and use this information to make the next move. This is consistent with psychological research showing that, when sequential dependencies exist, people can detect and exploit them (e.g., Estes, 1972).
Using this model, they found the following results: (a) the interaction between two agents of this type produced the seeming randomness; (b) the sequential patterns that were produced by this process were temporary and short lived; and (c) human subjects played similarly to a lag 2 network that was punished for ties: that is, people were able to predict their opponent’s moves by using information from the previous two moves, and people treated ties as losses.
For other work that attempts to make game theory more cognitively–psychologically realistic, see, for example, Axelrod (1984), Camerer (1997), and Juvina, Lebiere, and Gonzalez (2015), among others.
A Cognitive Simulation of Organizations
Another example is a simulation of organizations conducted based on the Clarion cognitive architecture (Sun, 2002, 2016), which helped to shed light on the role of cognition in organizations (Sun & Naveh, 2004).
The simulation focused on a typical task faced by organizations—classification (Carley, Prietula, & Lin, 1998). In this simulation, no one single agent has access to all the information relevant to making a decision, and separate decisions made by different agents are integrated. Organizational structures include two types: (a) teams, which treat individual decisions as votes, and the organization decision is the majority decision; and (b) hierarchies, in which the decision of a superior is based solely on the recommendations of subordinates. Information is accessible to each agent in two different ways: (a) distributed access, in which each agent sees a different subset of attributes, and (b) blocked access, in which several agents see exactly the same subset of attributes.
Because the Clarion cognitive architecture is intended for capturing all the essential cognitive-psychological processes (Sun, 2002, 2016), its cognitive parameters include, for example, learning rate, generalization threshold, probability of using implicit versus explicit processing, and so on. With these parameters, the results of the simulation closely accord with the patterns of the human data (e.g., with teams outperforming hierarchies, and distributed access being superior to blocked access; cf. Carley et al., 1998), far better than previous simulations, which shows the advantage of cognitive social simulation.
But what happens when cognitive parameters are varied? The statistical results show the superiority of team and distributed information access early on, and the disappearance or reversal of these advantages later. The analysis showed that the above trend did not depend on any one setting of parameters.
In sum, the cognitive social simulation with the Clarion cognitive architecture more accurately captured organizational performance data and led to deeper explanations for the results (see Sun & Naveh, 2004, for details). Furthermore, with Clarion, one can vary parameters that correspond to cognitive processes and test their effects on collective performance; thus, this approach may be used to predict human performance in organizational settings and to prescribe optimal or near-optimal cognitive abilities for individuals for specific tasks and organizational structures.
For other possibilities of cognitive social simulation of organization and group, see, for example, Carley et al. (1998); Clancey, Sierhuis, Damer, and Brodsky (2006); Clancey, Linde, Seah, and Shafto (2013); Helmhout (2006); Prietula, Carley, and Gasser (1998); and so on.
Some Other Cognitive Social Simulations
In addition, some cognitive social simulations may include evolutionary processes, for example, evolutionary simulation of social survival strategies (Cecconi & Parisi, 1998; Sun & Naveh, 2007), evolution of individual motivational processes (Sun & Fleischer, 2012), and simulating other issues relevant to evolution of cognitive processes in social settings (Kenrick, Li, & Butner, 2003; Kluver, Malecki, Schmidt, & Stoica, 2003).
Other cognitive social simulations include models of individual and collective motivation (e.g., Clancey et al., 2006; Wilson, Sun, & Mathews, 2009), personality and personality interaction (e.g., Quek & Moskowitz, 2007; Sun & Wilson, 2014), emotion and emotion contagion (e.g., Allen & Sun, 2016; Thagard & Kroon, 2006), and human morality (Bretz & Sun, 2018). For example, unified models of motivation, emotion, personality, moral judgment, and so on have been developed within the Clarion cognitive architecture (Sun, 2016), for the sake of in-depth understanding of these aspects together. For other models of emotions in social settings, see also Erisen, Lodge, and Taber (2014); Gratch, Mao, and Marsella (2006); and so on. These models further strengthen cognitive social simulation and its abilities to tackle deeper psychological issues involved in social processes. They help with not only the better understanding of motivation, emotion, personality, and so on, but also the better understanding of their roles in social interaction.
Furthermore, simulations of political behavior have been undertaken. For example, the Clarion cognitive architecture was applied to studying voter decisions in an election campaign (Schreiber, 2004). The ACT-R cognitive architecture was applied to produce a computational model of political attitudes, incorporating psychological theories with findings from electoral behavior (Kim et al., 2010).
An analysis of social issues based on some existing models of cognitive agents was proposed by White (in press). It was suggested that these models can help understand weighty social and political issues (such as those involved in international geopolitics), and they may lead to reasonable solutions of these consequential issues. Using cognitive social simulation to tackle these issues is an important suggestion.
Issues addressed by social simulation, especially cognitive social simulation, have been diverse. They include, for example, emotional interaction, crowd behavior, tribal customs, belief systems, academic publishing and citation, game playing, stock market dynamics, social cooperation, group interaction, organizational decision making, political behavior, evolution of language, formation of social norms, and countless others (see, for example, Sun, 2006).
Applications and Further Developments
By incorporating detailed cognitive models, one can take into consideration human cognition-psychology when predicting or explaining collective social outcomes (Sun, 2012; Sun & Naveh, 2004). Conversely, one can also take into consideration sociocultural processes in understanding individual mind (Nisbett, Peng, Choi, & Norenzayan, 2001; Vygotsky, 1962; Zerubavel, 1997). The result is better, more detailed, and more accurate models and simulations.
Cognitive social simulation is still at an early stage of development, given the relatively recent emergence of the two fields on which it is based (social simulation and cognitive modeling, including cognitive architectures). Many research issues and challenges remain to be addressed to better serve policy makers.
First, whether or not to use detailed cognitive models in social simulation is a decision that has to be made on a case-by-case basis. There are many reasons for using or not using detailed cognitive models. The reasons for using detailed cognitive models include the following: (a) cognitive realism may lead to more accurately capturing human data in social simulation; (b) with cognitive realism, one will be able to formulate deeper explanations for results observed, by basing explanations on cognitive factors rather than arbitrary assumptions; and (c) with detailed cognitive models, one can vary parameters that correspond to cognitive processes and test their effects on outcomes, and in this way, simulations may be used to predict outcomes based on cognitive factors or to improve performance by prescribing optimal cognitive abilities for specific tasks.
The reasons for not using detailed cognitive models in social simulation include the following: (a) it is sometimes possible to describe causal relationships at higher levels without referring to relationships at lower levels (Goldstone & Janssen, 2005); (b) complexity may make it difficult to interpret results in terms of their precise contributing factors; and (c) complexity also leads to longer running times and hence raises issues of scalability.
Another issue facing cognitive social simulation is validation of simulation results, including validation of cognitive models as part of social simulation. Validation of complex simulation models is always difficult (Axtell, Axelrod, & Cohen, 1996; Moss, 2006; Pew & Mavor, 1998). However, in this regard, adopting existing cognitive models as part of a cognitive social simulation may be beneficial: If one adopts a well-established cognitive model (a cognitive architecture in particular), the prior validation of that cognitive model may be leveraged in validating the overall simulation results. Therefore, there is a significant advantage in adopting an existing cognitive model.
This area of research will come to fruition in relation to better understanding cognition and sociality as well as their interaction (Sun 2006, 2012). Consequently, in terms of its relevance to policy making, we may examine briefly a few example cases below.
For instance, computational models of organizational structures and dynamics on the basis of cognitive models (as discussed earlier) can be useful to understanding or even designing organizational structures and makeups for improving organizational performance in various situations. Cognitive architectures have been applied to the simulation of organizational decision making (as described earlier; Carley et al., 1998; Sun & Naveh, 2004). Relatedly, there have also been cognitively based models of group dynamics (Clancey et al., 2006). These models can lead to significant applications in organizations of various types.
Industrial/organizational psychology needs to understand not only how goal setting, feedback, self-efficacy, and other parameters affect individual performance (Locke & Latham, 2013) but also how these parameters interact with social environments (e.g., team goals, supportive colleagues, emotion contagion, etc.) in affecting overall performance. High-fidelity cognitive social simulation can provide valuable information concerning interactions of these parameters and thus is useful in understanding implications of organizational practices and policies.
Ongoing work on computationally modeling emotion, motivation, personality, and other socially relevant psychological aspects may be useful to cognitive social simulation in terms of leading to applications. These models are useful not only for understanding motivation, emotion, and personality per se but also for designing relevant social mechanisms for channeling them for public good. For example, emotion contagion is prevalent in social settings; it may be useful for law enforcement to be able to anticipate crowd behavior in volatile situations (Parunak, Brooks, Brueckner, Gupta, & Li, 2014) in part based on modeling emotion contagion among a crowd.
Computational models of politics on the basis of individual cognition (as mentioned earlier) lead to detailed simulations of voter behavior, political opinion formation, emotional response, and emotionally colored political reasoning. These models can be useful tools for political mechanism design and for deciding on political strategies and coalition formation.
Other research directions involving cognitive modeling and social simulation are currently being actively pursued, including, for example, robot teaming (to generate useful social behavior among a group of robots) and battlefield simulation (with detailed cognitive models of agents; Pew & Mavor, 1998). Some of these research directions may lead to significant applications as well as making of relevant policies.
Overall, many directions of research pursued in cognitive social simulation have significant implications for understanding and making public and other relevant policies. These directions may lead to better, more cognitively and socially realistic simulations that address both fundamental theoretical issues facing social and cognitive scientists and practical policy matters facing policy makers.
Summary
This article surveys cognitive social simulation, which is at the intersection of cognitive modeling and social simulation. By integrating cognitive and social models, cognitive social simulation can address issues concerning both cognition-psychology and sociality. More importantly, cognitive social simulation can find important practical applications in relation to public and other policies in many areas. The present article may be considered an appeal to better utilize (and to further develop) cognitive social simulation for improved policy making.
Footnotes
Author’s Note
The cognitive models mentioned, including Clarion and ACT-R, are academic research programs, not commercial products.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This article was written while the author was supported (in part) by ARI Grant W911NF-17-1-0236.