Levels of Analysis in Computational Social Science

Peter M. Krafft ([email protected])

Thomas L. Griffiths (tom [email protected])
Department of Psychology, University of California, Berkeley
Social Science Matrix, University of California, Berkeley

Abstract level of analysis—representing social systems as distributed

Marr’s levels of analysis constitute one influential approach to computer programs. The naı̈ve application does not neces-
the central program of cognitive science—the multilevel anal- sarily extend through to the computational level, in which the
ysis of cognition as information processing. The distinctive combination of agent behavior, taken holistically, would have
aspects of Marr’s framework are an emphasis on identifying
the computational problems and constraints faced in cognition, to yield coherent distributed computation at the population
and conceptual machinery to relate cognitive mechanisms to level. To address this difficulty, we pursue a program of iden-
that computational level of analysis. Although related ideas tifying which social behaviors and structures can be produc-
have been explored in a range of social science disciplines,
Marr’s framework, and particularly its notion of the precise tively conceptualized as having computational roles.
formulation of computational problems and solutions, has yet There are several existing approaches in the social sciences
to be applied widely in social analysis. In the present work related to Marr’s framework. The field of organization sci-
we develop a formulation of Marr’s levels for social systems,
provide examples of this approach, and address potential criti- ence has explicitly adopted information processing perspec-
cisms. The consequence is a computational perspective on the tives since the seminal work of Herbert Simon. Simon, who
sociological school of structural functionalism, and an appara- was an early thinker on the topic of information processing
tus for conducting multiscale analysis of social systems.
in the context of human behavior (Simon, 1978), also applied
Keywords: Computational social science; Marr’s levels of
analysis; structural functionalism; analytical sociology; com- these ideas to organizations. In one famous passage, Simon
putational social theory writes: “In the post-industrial society, the central problem
is not how to organize to produce efficiently ... but how to
Introduction organize to make decisions, that is, to process information”
Marr (1982) famously argued that any information processing (Simon, 1973, p. 269–270). Financial markets are also com-
system can be analyzed at three levels, that of (1) the compu- monly understood as information processing systems. In a
tational problem the system is solving; (2) the algorithm the classic economics paper, Hayek states: “the economic prob-
system uses to solve that problem; and (3) how that algorithm lem of society ... is a problem of the utilization of knowledge
is implemented in the “physical hardware” of the system. not given to anyone in its totality” (Hayek, 1945, p. 519–520).
This decomposition offers both functional and mechanistic Marr’s algorithmic and implementation levels are akin to
perspectives on information processing systems. Marr’s main mechanistic explanations. Mechanistic explanations are pop-
aim was understanding psychology and the human brain, and ular in the classical area of mathematical sociology, such as
his levels of analysis have proven to be a useful conceptual in Schelling’s seggregation model (Schelling, 1971) or Gra-
tool for generations of cognitive scientists after him. novetter’s threshold model (Granovetter, 1978), and remain
Considering that aspects of human cognition can be pro- popular in the modern area of analytical sociology (Hedström
ductively viewed as information processing, and that social & Bearman, 2009). Recent progress in the field of economics
groups consist in part of sets of people who exchange infor- and computation highlights the algorithmic side of the math-
mation, it is natural think about how Marr’s levels of analysis ematical notion of game-theoretic equilibrium (Daskalakis,
could be productively applied in the analysis of social sys- Goldberg, & Papadimitriou, 2009). Mechanisms have also
tems. An immediate difficulty occurs with a naı̈ve applica- been a target of inquiry in organization science, such as in the
tion, however. Taking the disciplinary commitments of cog- study of transactive memory (Wegner, 1987).
nitive science for granted, we can clearly model social sys- These existing lenses in the social sciences fit naturally
tems as distributed computer programs. Yet, there is no guar- within Marr’s framework, and therefore point towards a syn-
antee at all that the resulting computer programs implement thesis of a cognitive, information-processing view of a wide
any coherent distributed solution to particular computational variety of social systems. At the same time, many of these
problems. Far from all human collective behavior, or for classic works did not draw explicit parallels to distributed
that matter far from all combinations of synthetic intelligent computation, or did not leverage the hierarchies of abstrac-
agent behavior, has any functional purpose. Well-recognized tion familiar to computer scientists that Marr deploys. Cog-
examples of collective dysfunction resulting even from in- nitive scientists have begun to explicitly explore the appli-
telligent agents include information cascades (Bikhchandani, cation of Marr’s levels to social systems. Hutchins (1995)
Hirshleifer, & Welch, 1992), phantom traffic jams (Kerner & pioneered the application of Marr’s levels to social systems
Konhäuser, 1993), and the tragedy of the commons (Ostrom, in his ethnography of distributed cognition in team behavior
2015). The naı̈ve application of Marr’s levels of analysis to on a naval vessel. To Hutchins, the computation performed
social systems therefore only extends as far as an algorithmic by a naval vessel was that of navigation—calculating where
you are and determining how to get where you want to go.
Hutchins provided a detailed account of how this function
is accomplished by the crew members and their interactions
with each other and with artifacts on the ship.
Hutchins’ example makes it clear that the explicit applica-
tion of Marr’s levels can be productive in the context of teams
and organizations. However, one of the reasons this exam-
ple easily fits into Marr’s framework is because teams and
organizations have well-defined group boundaries and have
functions that are explicit in the goals of these groups. These
goals then dictate the information processing challenges the
groups face. An important outstanding question is to what
extent Marr’s approach can be applied to more loosely orga-
nized social systems that are often the subject of sociology.
The functionalist lens, used in sociology for a variety
of less strictly organized and less explicitly engineered so-
cial systems, provides reason to believe that there is space
for such an attempt to be fruitful. Structural or sociologi-
cal functionalism—i.e., the pursuit of understanding social
structures and behavior in terms of how they solve social
problems—is one of the classic theoretical perspectives in so-
ciology. Many early sociologists held views that a variety of
social phenomena played functional roles in society. For in-
stance, Spencer (1898) advocated for an equilibrium view of
society and drew extensive analogies between social and bi- Figure 1: Marr’s levels of analysis for waiting in line.
ological function. Durkheim (1893) presented a functionalist
argument that division of labor acts as a mechanism of so-
cial solidarity promoting a cohesive social bond. Although each standing behind another. This social behavior imple-
functionalism was and continues to be controversial in soci- ments the computation of a first-in, first-out (FIFO) queue. A
ology (Weber, 1922; Giddens, 1984), scholars still lean on it FIFO queue is a simple function used in computer science,
in modern studies. Yet, unlike in organization science, ex- for example to prioritize computer processes in the CPU. A
plicit information processing analogies are barely ever used FIFO queue takes as input a stream of entries, maintains the
in sociology. order of those entries, and outputs entries in that order. The
In the present work, we explicate the application of Marr’s representation that is used to solve the FIFO queue problem
levels to loosely organized social systems, review examples in the case of waiting in line is to maintain a linked list data
of recent work that fit within this paradigm, address chal- structure between elements of the queue, and pop elements
lenges to this approach, and explore its potential and limi- off the list as needed. A linked list is another data struc-
tations. The main benefit of Marr’s approach is that com- ture used computer science, in which each entry contains a
putation provides an expressive language for high-level, ab- “pointer” to the next element in the list. To “pop” a linked list
stract theory, while providing the conceptual machinery to re- means to remove the head element. In the example of waiting
late that abstract level to mechanistic explanations. Computa- in line, the distributed algorithm that implements this linked
tional social theory can therefore be precisely specified, and list is for each person in the line to keep track of who is ahead
tested via its relation to algorithmic and behavioral descrip- of them. The physical implementation used to keep track of
tions. At the same time, Marr’s charge to identify compu- who is ahead of you in the line is simply to stand behind that
tational problems that information processing systems solve person. Figure 1 illustrates this example.
could provide inspiration for research questions in computa- There are many failure modes to standing in line. Two peo-
tional social science. ple can arrive at a similar time or be in a similar position and
not be certain who is ahead of whom. Some people cut in
Illustrative Examples line. Sometimes lines fail to form at all or totally collapse
Before more carefully defining Marr’s levels for social anal- and become disorganized crowds of people waiting. There
ysis, we begin with three motivating illustrative examples. is also cultural variation in how much importance or value
people place on lines as a useful mechanism. Taking Marr’s
Waiting in Line approach abstracts away these details and exposes the under-
A simple example that illustrates Marr’s three levels of anal- lying information processing challenge at the heart of stand-
ysis in a social system is waiting in line. A line, for instance ing in line. Our ability to reason about the computational
outside a professor’s office, consists of a group of people, function of waiting in line can also suggest other engineered
solutions. Waiting in line is not the only solution to the FIFO computational problem. The computational problems in our
queue problem. For example, some delis, grocery stores, and examples were implementing a FIFO queue, resource alloca-
government offices implement ticket and announcement sys- tion, and distributed inference. Other common computational
tems that obviate the need to keep your own place in line. problems in social systems include aligning group mem-
ber preferences and solving coordination problems (Krafft,
Status Hierarchies 2018). Unlike in cognition, in which computational problems
As another example, many species of animals maintain some are frequently posed by the external environment, many com-
kind of social hierarchy, influencing the interactions between putational problems faced by groups are endogenous. The
animals in a group and the way that they allocate resources need to coordinate is one example. The need to coordinate
such as food. At a computational level, this structure can be is an inherent result of existing as differentiated people. An-
viewed as a solution to the problem of performing resource other type of endogeneity is in problems that are created by
allocation with a minimum of conflict—an alternative to a history dependence. For instance, Durkheim offers that one
costly free-for-all whenever resources become available. At view of the division of labor could be that by increasing our
the algorithmic level, there are many kinds of algorithms that ability to create goods to relieve our increasing fatigue, divi-
can be used to impose an ordering on a group through pair- sion of labor functions in part to meet the needs created by its
wise comparison (i.e. fights or displays)—a whole branch of very existence (Durkheim, 1893).
theoretical computer science is devoted to questions about The computational level of analysis is important because it
sorting. However, here the implementation details matter: allows the researcher to answer “why” questions—to under-
most sorting algorithms involve maintaining some kind of stand why people behave in a certain way. In order to justify
global record of the current ordering of the items being sorted. a teleological interpretation of a social function, that function
In an animal group, each animal needs to maintain such a should either be explicitly intended or otherwise be evidently
record independently. Thus it makes sense to think about addressing a problem that threatens the group. For instance,
distributed algorithms in which individual animals operate as we can say that conventions about which side of the street to
computational elements. Flack and Krakauer (2011) applied drive on exist in order to solve a coordination problem. Some
exactly this approach to modeling the decisions of monkeys to coherent distributed computations do not meet this criterion
engage in fights, viewing these individual decisions as form- of solving a computational problem associated with an inten-
ing “adaptive social circuits”. tion or a need of the group. In the tragedy of the commons,
rational agents are computing an equilibrium, and therefore
Rumors and Collective Sensemaking accomplishing a computational function, but this outcome is
As a final motivating example, a classical and now well- neither intended nor meeting a need. Therefore this collective
supported sociological theory of rumors conceptualizes ru- behavior cannot be productively interpreted as functional, and
mors as a natural part of a process of collective sensemaking Marr’s computational level does not apply.
(Shibutani, 1966; Bordia & DiFonzo, 2004; Huang, Starbird, Another important qualification in the social case is that
Orand, Stanek, & Pedersen, 2015). According to this view, the computational problem is one faced by the group, com-
people try to make sense of the world together when they munity, or society. Every individual in a group has their own
find themselves in uncertain environments. A computational problems and goals, and some behavior will be oriented to-
perspective of rumors based on this view is as functioning to wards those individual needs and not any shared needs of the
communicate hypotheses about the state of the environment group. Selfish behavior of this sort is one reason why we can-
(Krafft, Zhou, Edwards, Starbird, & Spiro, 2017). This per- not treat all compositions of rational behavior as functional
spective frames rumors as oriented towards a distributed in- group behavior. This issue is at play in the tragedy of the
ference problem of inferring the state of the world given the commons and other social dilemmas from game theory.
evidence at hand. The definition of social functions is also only with respect
to the boundaries of the group being analyzed, and does not
Marr’s Levels for Social Systems represent a moral judgment. Accomplishing a computational
Having motivated computational and algorithmic perspec- function in one group can cause problems for other groups;
tives of social systems through our examples, we now expand consider the case of one group finding a new place to build
upon the usage of Marr’s levels of analysis in social systems. a settlement and displacing another group. In line with We-
ber’s interpretive approach (Weber, 1922), insofar as we are
Computational Level aiming to understand why people are engaging in certain so-
The first level of analysis Marr defined is the computational cial behaviors, we must interpret function with respect to the
level. The computational level describes the problem that an values of the people in the group being analyzed.
information processing system is oriented towards solving.
The information processing function that the social system Algorithmic Level
accomplishes may be explicit due to design or implicit, as Marr’s second level of analysis is the algorithmic level. The
in Merton’s manifest versus latent functions (Merton, 1949). algorithmic level describes the way in which a computational
For this level of analysis to apply, the group must face some problem associated with an information processing system is
solved. An algorithm involves both the representations of in- coordination, there is a wealth of literature in computer sci-
formation used and the transformations of those representa- ence on engineering distributed systems.
tions. In social systems, the fundamental algorithms at play A second benefit is a rigorous approach to providing math-
are most readily conceptualized as distributed algorithms, in ematical evidence for functionalist sociological theories. Tak-
which multiple people are participating as agents akin to net- ing the rumor example again, suppose we wanted to provide
worked computer processors. This perspective of social pro- evidence that rumors function as a mechanism of collective
cesses as distributed algorithms is closely related to agent- sensemaking. Suppose we can show that a distributed in-
based modeling (Macy & Willer, 2002), the study of social ference algorithm as a behavioral model explains observed
mechanisms in analytical sociology (Hedström & Bearman, behavior better than alternative mechanisms, such as a con-
2009), and the study of natural algorithms in theoretical com- tagion model or a thermodynamic model. The evidence for
puter science (Chazelle, 2009). that distributed algorithm then in turn provides evidence for
The critical criterion for an algorithmic explanation within the functional interpretation of rumors since there is a math-
Marr’s framework is that the behavior being examined offers ematical relationship between the algorithmic model and the
a proper solution to the computational problem posed in the problem of distributed inference.
computational level of analysis. This criterion pushes beyond A final benefit is for design. Once a social problem is spec-
purely descriptive studies of social mechanisms, as in many ified precisely as a computational problem, then we can do
agent-based or rational models, towards a formal relationship more than just understand how current social behavior might
between mechanisms and social functions. In the example of address this problem. We can also search for alternative social
waiting in line, the FIFO queue is accomplished if each per- behaviors or structures that better solve the problem accord-
son keeps their place. While deviant behavior such as line- ing to some criteria. The value of precise computational spec-
cutting could be included in an agent-based model, a strict al- ification is that this search through design space can be auto-
gorithmic analysis does not accommodate cases when some mated. An example of this approach is in automated mecha-
people cut in line for no reason other than their own self- nism design (Conitzer & Sandholm, 2003).
interest, because this behavior undermines the correct com-
putation of the FIFO queue. Challenges to the Approach
Implementation Level There are several interrelated challenges and potential cri-
tiques of the indiscriminate application of Marr’s approach
The final level of analysis Marr defines is the implementation to social systems. We now address what we view as the
level. An algorithm is an abstract process-level description. major challenges. Our responses to these challenges center
The implementation level explains how the algorithm is im- around an argument that a program of Marr’s approach to
plemented through actual interaction of basic elements. In social systems aims to produce useful, idealized hierarchical
social systems, the definition of the implementation level is mathematical descriptions, but should not be conducted with-
contingent on what elements of the system are taken as prim- out also paying careful attention to the specifics of the social
itive. Typically, social systems are reduced to psychologi- context being studied and the political aspects of that inquiry.
cal processes, and connections to neural processes are left to
cognitive scientists. This division leaves the implementation Multiagent Systems versus Human Social Systems
level to be concerned with psychological processes, details
One potential criticism that we can readily dismiss involves
of social interaction, and contextual elements such as geogra-
the difference between human social systems and artificial
phy, social network structure, and artifacts in the environment
multiagent systems or distributed computer systems. Com-
as building blocks. The implementation level can be thought
puter networks offer quite different affordances and con-
of as a second, lower-level algorithmic analysis.
straints as compared to social systems. For instance, comput-
ers can easily communicate their entire internal states with
Benefits of the Approach complete precision to each other. Communication is much
There are several potential gains to be had from employing harder for people, but at the same time, people have a richer
Marr’s levels of analysis to understand social systems. One range of distinctive forms of communication, including sym-
benefit is a deductive approach to discovering mechanisms. bolic and cultural systems. Social networks, the physical-
In analytical sociology, the discovery of mathematical de- ity of human interaction, social norms and institutions, and
scriptions for social mechanisms is often post hoc and in- many other contextual factors form additional components
ductive from observations. Marr’s levels provide a deductive, that must be considered in the case of social systems. Al-
reverse-engineering approach. In this approach, the compu- though distributed computer systems and distributed social
tational problem being faced by a group is specified first, and systems clearly have widely differing constraints and affor-
then algorithms to solve that problem are explored. For in- dances, the mathematical language we use to describe both
stance, in the case of conceptualizing rumors as distributed types of systems, the classes of algorithms that are employed,
inference, we can look to the literature on algorithms for dis- and some fraction of the computational problems each type
tributed inference in search of mechanisms. To understand of system faces could still be similar.
Methodological Individualism At the same time, we might think that segregation is actually
a maladaptive characteristic of a population. In a more recent
A classic criticism of functionalism is that of methodological
example, a group of analytical sociologists presented a mul-
individualists or “reductionists”. The view of methodologi-
tiscale analysis of adolescent sexual behavior, and showed
cal individualism would assert that a group-level functional-
that the behavioral mechanisms of these people led to net-
ist perspective is unnecessary for explaining the behavior of
works that tended to have structures similar to spanning trees
social systems. Under such a view, any group-level structure
(Bearman, Moody, & Stovel, 2004). Spanning trees are good
supervenes on the individual-level beliefs, intentions, plans,
for the sexual health of the community in some ways but bad
and behaviors. A social system therefore cannot be properly
in other ways, and thus do not serve a clear function. These
understood as having a function. To have a function means
analyses benefit from the same mathematical machinery that
that the social system as a whole has a causal role in a broader
we use in the computational level analysis, but fall outside its
ecosystem. But according to the reductionist view, the social
scope in our definition.
system as a whole plays no causal role in the system dynam-
ics. The only causally relevant entities are the components. Dysfunctional Collective Behavior
There are several responses to such a criticism. The re-
sponse requiring the weakest logical commitments is that Another concern is individual behaviors that lead to incoher-
function can serve as a useful description that succinctly ent or unstructured collective behavior, and individual or col-
summarizes the behavior of the system, without making any lective behaviors that appear functionally oriented but fail to
causal claims. The usefulness of the description alone justi- accomplish any function. Social behavior that leads to in-
fies a functionalist inquiry. When conceptualizing social sys- coherent or unstructured collective behavior, such as people
tems “as if” they had functions allows us to better understand going about their own individual business within their homes,
them, then such concepts are valuable. Another response is to simply may not have group-level structure that lends itself to
assert that there can be multiple scales of causal explanations, illuminating interpretation via Marr’s levels. One prominent
and functions serve a causal role at an aggregate level. For in- and perhaps surprising example of this sort is agent-based
stance, in a counterfactual world where people have no way models. Although agent-based models can be described as
to implement a FIFO queue at the professor’s office, then the computer programs, their aggregate dynamics are sometimes
operations of the group—who gets in when—would be fun- chaotic or unstructured. The scope of Marr’s levels therefore
damentally different. Therefore the operation of office hours is not as wide as the class of all processes that can be de-
can rightfully be conceptualized as having a dependence on scribed as distributed computer programs. Other behaviors
the ability of people to implement a FIFO queue. may appear functionally oriented but are suboptimal or to-
A related response is that certain functions are irreducible tally dysfunctional. Here, Marr’s approach simply may not
emergent properties of the social system, meaning that the apply if the social system does not have group-level informa-
behavior of the social system cannot be properly understood tion processing characteristics.
without understanding its emergent functions. This argument
asserts that there is downward causality from emergent func- Sociological Critiques
tion to the constituents of the system. For instance, consider A final set of threats to applying Marr’s levels to social sys-
two competing hunting groups whose members must hunt tems are inherited from other challenges to sociological func-
together to be successful, and suppose the members of one tionalism. Despite still being influential in contemporary so-
group have synchronized clocks. It is the ability of the group ciology, functionalism has been criticized for abstracting and
to accomplish the task of coordination that allows the group obscuring many key details of social phenomena. Conflict
to be more effective, not anything about the individuals in and theorists have emphasized how functionalism diminishes the
of themselves. If the clocks did not accomplish the function struggle of marginalized groups, the importance of revolu-
of coordination, the individuals would not benefit. tionary change, and the role of individual human agency in
society. In a somewhat separate line of critique, Giddins’
Non-adaptive Functions structuration theory explores how function and structure co-
We argued in our definition of the computational level of anal- evolve continuously across space and time and cannot be
ysis that the computational functions being analyzed must be neatly separated (Giddens, 1984).
beneficial to the group in order for the function to have a tele- We follow Weber (1922) in responding to these criticisms
ological interpretation, which is an implicit aim of Marr’s by noting firstly that a functional description can still be use-
computational level. Some collective behavior implements ful for certain ends, although non-functional aspects must
coherent computation that is either harmful or epiphenom- be considered for a complete treatment of any system; and
enal. For instance, Schelling’s (1971) segregation model a functionalist analysis may still be illuminating to see how
showed how small individual biases could lead to a popu- a population deviates from an idealized solution. That said,
lation clustering itself according to attributes such as race. we must always keep in mind the balance between the clarity
Clustering or sorting could then be said to be an information provided by abstraction and the frequent importance of the
processing function implemented by Schelling’s mechanism. details that are abstracted away.
 Conclusion Bordia, P., & DiFonzo, N. (2004). Problem solving in social inter-
actions on the internet: Rumor as social cognition. Social Psy-
Analysis of social systems is challenging in part due to the chology Quarterly, 67(1), 33–49.
diversity and complexity of the people in these systems and Chazelle, B. (2009). Natural algorithms. Proceedings of the Twenti-
eth Annual ACM-SIAM Symposium on Discrete Algorithms, 422–
their interactions. Abstraction of social processes should be 431.
approached with caution but can help to highlight generaliz- Conitzer, V., & Sandholm, T. (2003). Applications of automated
able insights and underlying principles at play. In the present mechanism design. UAI-03 Workshop on Bayesian Modeling Ap-
plications.
work we have outlined the application of Marr’s levels of Daskalakis, C., Goldberg, P. W., & Papadimitriou, C. H. (2009).
analysis to a broad range of social systems. This framework The complexity of computing a Nash equilibrium. SIAM Journal
provides machinery for abstraction, and for relating abstract on Computing, 39(1), 195–259.
Durkheim, E. (1893). Division of labor in society. Paris: Alcan.
levels to lower levels of explanation. Flack, J. C., & Krakauer, D. C. (2011). Challenges for complex-
Given our discussion of criticisms in the previous section, ity measures: A perspective from social dynamics and collective
social computation. Chaos: An Interdisciplinary Journal of Non-
we can make an informed attempt at outlining the scope of linear Science, 21(3), 037108.
Marr’s approach for social systems. Marr’s levels have clear Giddens, A. (1984). The constitution of society: Outline of the
utility and have been used in cases where groups have an theory of structuration. University of California Press.
Granovetter, M. (1978). Threshold models of collective behavior.
explicit shared goal of executing an information processing American Journal of Sociology, 83(6), 1420–1443.
task, such as in teams and organizations. We have argued that Grosz, B. J., Hunsberger, L., & Kraus, S. (1999). Planning and
Marr’s levels of analysis can also be useful in cases where acting together. AI Magazine, 20(4), 23.
Hayek, F. A. (1945). The use of knowledge in society. The American
people are self-organized—whether by coincidence, by inten- Economic Review, 35(4), 519–530.
tion, or through biological or cultural evolution—to execute Hedström, P., & Bearman, P. (2009). The Oxford handbook of ana-
coherent information processing. Cases that are less appro- lytical sociology. Oxford University Press.
Huang, Y. L., Starbird, K., Orand, M., Stanek, S. A., & Pedersen,
priate are when group behavior is unstructured, incoherent, or H. T. (2015). Connected through crisis: Emotional proximity
structured but not oriented towards information processing. and the spread of misinformation online. Proceedings of the 18th
There are several classes of computational problems ACM Conference on Computer Supported Cooperative Work &
Social Computing, 969–980.
and algorithms that may be useful in deploying Marr’s Hutchins, E. (1995). Cognition in the Wild. MIT Press.
levels of analysis for social systems. Multiagent systems is Kerner, B. S., & Konhäuser, P. (1993). Cluster effect in initially
one of the most relevant areas (Shoham & Leyton-Brown, homogeneous traffic flow. Physical Review E, 48(4), R2335.
Krafft, P. (2018). A simple computational theory of general col-
2008). The area of multiagent systems provides a wide lective intelligence. Topics in Cognitive Science (topiCS) Special
variety of formalisms that are useful for both specifying Issue on Computational Approaches to Social Cognition.
problems and algorithms, including multiagent decision Krafft, P., Zhou, K., Edwards, I., Starbird, K., & Spiro, E. S. (2017).
Centralized, parallel, and distributed information processing dur-
problems (Bernstein, Zilberstein, & Immerman, 2000) and ing collective sensemaking. Proceedings of the 2017 CHI Con-
computational models of shared cooperative activity (Grosz, ference on Human Factors in Computing Systems, 2976–2987.
Hunsberger, & Kraus, 1999). There are also many distributed Lynch, N. A. (1996). Distributed algorithms. Morgan Kaufmann.
Macy, M. W., & Willer, R. (2002). From factors to actors: Compu-
algorithms outside the literature on multiagent systems, such tational sociology and agent-based modeling. Annual Review of
as fault-tolerant distributed algorithms for consensus (Lynch, Sociology, 28(1), 143–166.
1996). Distributed machine learning is another promising Marr, D. (1982). Vision: A computational approach. MIT Press.
Merton, R. K. (1949). Social theory and social structure. The Free
area to draw upon. Future work could deploy Marr’s levels Press.
of analysis to further explore links between developments in Ostrom, E. (2015). Governing the commons. Cambridge University
these areas and the study of social systems. Press.
Schelling, T. C. (1971). Dynamic models of segregation. Journal
of Mathematical Sociology, 1(2), 143–186.
Acknowledgments. Special thanks to Josh Tenenbaum, Sandy Shibutani, T. (1966). Improvised news: A sociological study of
Pentland, David Krakauer, and Jessica Flack who helped stimulate rumor. The Bobbs-Merrill Company.
Shoham, Y., & Leyton-Brown, K. (2008). Multiagent systems: Al-
an initial investigation into these ideas; to Josh for his suggestion to gorithmic, game-theoretic, and logical foundations. Cambridge
think about the example of waiting in line; and to Nick Logler, Steph University Press.
Ballard, and Lavi Aulck for their special assistance. This work was Simon, H. A. (1973). Applying information technology to organi-
zation design. Public Administration Review, 33(3), 268–278.
supported by DARPA Cooperative Agreement D17AC00004. Simon, H. A. (1978). Information-processing theory of human prob-
lem solving. In W. K. Estes (Ed.), Handbook of learning and cog-
References nitive processes (Vol. 5, pp. 271–295). Hillsdale, NJ: Lawrence
Erlbaum Associates.
Bearman, P. S., Moody, J., & Stovel, K. (2004). Chains of affec- Spencer, H. (1898). The principles of sociology. New York: D.
tion: The structure of adolescent romantic and sexual networks. Appleton and Company.
American Journal of Sociology, 110(1), 44–91. Weber, M. (1922). Economy and society: An outline of interpretive
Bernstein, D. S., Zilberstein, S., & Immerman, N. (2000). The com- sociology (Vol. 1).
plexity of decentralized control of Markov decision processes. Wegner, D. M. (1987). Transactive memory: A contemporary anal-
Proceedings of the Sixteenth Conference on Uncertainty in Ar- ysis of the group mind. In Theories of group behavior (pp. 185–
tificial Intelligence, 32–37. 208). Springer.
Bikhchandani, S., Hirshleifer, D., & Welch, I. (1992). A theory of
fads, fashion, custom, and cultural change as informational cas-
cades. Journal of Political Economy, 100(5), 992–1026.