Self-Organized Fish Schools: An Examination of Emergent Properties

Author(s): Julia K. Parrish, Steven V. Viscido and Daniel Grünbaum

Source: Biological Bulletin, Vol. 202, No. 3 (Jun., 2002), pp. 296-305
Published by: Marine Biological Laboratory
Stable URL: http://www.jstor.org/stable/1543482 .
Accessed: 28/06/2014 08:34

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Marine Biological Laboratory is collaborating with JSTOR to digitize, preserve and extend access to
Biological Bulletin.

http://www.jstor.org

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions
Reference: Biol. Bull. 202: 296-305. (June 2002)

Self-OrganizedFish Schools: An Examination

of Emergent Properties

JULIA K. PARRISH1'2'*,STEVEN V. VISCIDO2, AND DANIEL GRUNBAUM3

1School of Aquatic and Fishery Sciences, Box 355020, Universityof Washington,Seattle, Washington
98195-5020; Zoology Department, Universityof Washington;and 3 School of Oceanography,
2

Universityof Washington

Abstract. Heterogeneous, "aggregated" patterns in the frameworkthat synthesizes much of this previouswork, and
spatial distributions of individuals are almost universal to identify relationshipsbetween behavioralparametersand
across living organisms,from bacteriato higher vertebrates. group-level statistics.
Whereas specific featuresof aggregationsare often visually
strikingto humaneyes, a heuristicanalysis based on human Types of Aggregations
vision is usually not sufficient to answer fundamentalques-
tions abouthow and why organismsaggregate.What are the Aggregation occurs in the smallest organisms-bacte-
individual-level behavioral traits that give rise to these ria-and the largest-whales-and spans virtually the en-
features? When qualitatively similar spatial patterns arise tire extant diversity of taxon, habitat, trophic level, life-
from purely physical mechanisms, are these patterns in history strategy, degree of mobility, and many other
organisms biologically significant, or are they simply epi- biological characteristics (Parrish and Edelstein-Keshet,
phenomena that are likely characteristics of any set of 1999; Camazine et al., 2001). Physical aggregationcan be
interactingautonomous individuals? If specific features of regardedas partof a continuumin groupintegration.At one
spatial aggregationsdo confer advantagesor disadvantages end of this continuumare territorialanimals with little need
in the fitness of group members,how has evolution operated to engage in information transfer and no need for group
to shape individualbehaviorin balancingcosts and benefits structure.At the other end are highly integrated,long-term
at the individualand group levels? Mathematicalmodels of associations between individuals that know-and perhaps
social behaviorssuch as schooling in fishes provide a prom- are even related to-other members of the group, and in
ising avenue to address some of these questions. However, which memberscan potentiallyhave high ratesof directand
the literatureon schooling models has lacked a common indirect information exchange. Honeybee hives, cetacean
framework to objectively and quantitatively characterize pods, and humancommunitiesare examples of these highly
relationshipsbetween individual-levelbehaviorsand group- integratedgroups (Wilson, 1975). In these systems, estab-
level patterns. In this paper, we briefly survey similarities lished pathways of long-term communication between
and differences in behavioral algorithms and aggregation known individuals (clones, siblings, reciprocating group
statistics among existing schooling models. We present members),or at predeterminedlocations (the hive, the calv-
preliminary results of our efforts to develop a modeling ing grounds, the dinner table), may supplementimmediate
sensory contact. Thus, the members remain part of the
* To whom correspondence should be addressed. E-mail: jparrish@ "group"even while they range widely in space. In these
u.washington.edu dispersedgroups,coordinatedfunctioncan be maintainedas
This paper was originally presented at a workshop titled The Limits to long as the necessary information is transferredbetween
Self-Organizationin Biological Systems.The workshop,which was held at group members (e.g., the sensory integration systems of
the J. Erik Jonsson Center of the National Academy of Sciences, Woods
Norris and Schilt, 1988; Schilt and Norris, 1997), though
Hole, Massachusetts,from 11-13 May 2001, was sponsoredby the Center
for Advanced Studies in the Space Life Sciences at the MarineBiological distance between interactingcomponents inevitably affects
Laboratory,and funded by the National Aeronautics and Space Adminis- the evolution and stability of emergent group properties
trationunder CooperativeAgreement NCC 2-896. (Hillier and Hanson, 1990; Latane et al., 1995).
296

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions
 SELF-ORGANIZEDFISH SCHOOLS 297

Between these extremes of group integration are what ena, and to more explicitly and mechanisticallydescribethe
could be considered "prototypical"animal aggregations- links between individual behaviors and group pattern.
herds, swarms, flocks, and schools. Within the fishes, over Dynamic patterns and movement are necessary charac-
50% of species school-that is, display synchronous and teristics of many biological aggregations, and are perhaps
coordinatedmovement-at some point in their life histories better criteria to distinguish adaptive responses from epi-
(Shaw, 1978), and an unknownadditionalnumberaggregate phenomena. For example, fish schools display complex
more coarsely. Prototypical aggregations exhibit coordi- emergent properties such as coordinated motion and di-
nated motion, but group members are generally unrelated rected activity. Compression,hourglass, vacuole, fountain,
and never develop lasting relationships(in the game-theo- and flash expansion (Fig. 2; Pitcher and Parrish, 1993) are
retic, "tit-for-tat"sense) with othermembers.Many of these all maneuversthat minimize predatoryrisk only if all mem-
groups are extremely large (e.g., a school of a billion bers performthem correctly (e.g., Parrish,1989). Parabolic
herring). Individuals in such groups interactwith a neigh- formationsin tuna schools may allow cooperative hunting
borhood of other members, but those may representa van- advantages (Partridgeet al., 1983). These emergent group
ishingly small fraction of the group as a whole. This sug- propertiesappearto have readily apparentbiological inter-
gests thatmechanismswhich maximize informationtransfer pretations, and physical analogs may be harder to find.
among individuals could be evolutionarily beneficial. One However, these dynamic group properties clearly confer
example of this is a repeatedarrangementwithin the group, more evolutionary advantages under some circumstances
reminiscentof crystal lattices, in which individuals assume than others (e.g., evading small predatorsthat target indi-
preferredpositions and orientationsrelative to their neigh- viduals while becoming targets for large predators that
bors. Such arrangementscould, for example, maximize sen- target groups; Parrish, 1993). Furthermore,the key evolu-
sory contact between members in such as way as to reflect tionaryquestion remains:do these behaviors involve trade-
ambient conditions, and the organisms' predominantsen- offs between short-termgain for the individual and long-
sory systems, morphology, etc. (Parrish,1992; Parrishand term functioning of the group? If so, what is the
Edelstein-Keshet,2000). evolutionarycontext in which selection for these behaviors
We can characterizepossible behavioral adaptationsin occurs?
members of these groups on at least two levels: (1) short-
term reactionsto modify position with respect to immediate Traffic Rules
neighbors; and (2) behavioralresponses that do not neces-
sarily improve position relative to immediate neighborsbut Assuming structure is advantageous, how is it main-
that contributeto group-level characteristicsthat ultimately tained? Laboratoryand field attemptsto address this ques-
benefit the individualby benefitingthe group. These group- tion in fish schools have been limited (Partridgeand Pitcher,
level adaptationsare among the most fascinating-and the 1980; Aoki et al., 1986; Parrishand Turchin, 1997), in part
most difficult to assess-aspects of animal aggregations. because obtainingthree-dimensionaltrajectorieson specific
individuals for a relevant period of time is difficult. Data
Pattern versus Function that do exist are typically from highly artificial conditions
(e.g., relatively small schools in highly lit still-watertanks;
Human perception tends to recognize attributes of the Parrish and Turchin, 1997). Three-dimensional tracking
whole: an even density profile, polarity, distinct edges, or techniques have not yet advanced to the stage where it is
specific shape. While it is temptingto assume that conspic- feasible to observe large schools (i.e., over 10), in three
uous features of biological aggregationsare somehow ben- dimensions, over long times (i.e., for more than seconds).
eficial, the existence of qualitatively similar patterns that Despite these difficulties, quantitativelyaccurate observa-
arise from physical phenomenashows that this need not be tions of fish behavior within schools will undoubtedlybe-
the case. For example, some shapes found in three-dimen- come available in the next few years. However, while those
sional schools (e.g., torus and funnel) are echoed in two- datawill providea means of assessing short-termbehavioral
dimensional insect configurations(e.g., wheel), suggesting responses by fish to neighborswithin a group, they will not
that such shapes may be adaptive for group members (Fig. by themselves provide the strong linkage between individ-
1). However, these patternscould be evolutionarilyneutral, ual and group characteristicsthat we requireto understand
or even pathological. Virtually identical shapes can be the mechanics and evolution of schooling behaviors.
found in a wide rangeof inanimateaggregations,from water Making this linkage requires an additional approach,
vapor to planets, in which these patternsarise from simple namely, mathematicalor computationalmodels of school-
abiotic interactionsbetween individual components, in the ing behavior.These models posit a specific, quantitativeset
absence of evolutionarydynamics. Key steps in understand- of behavioralinteractions-essentially, they create a set of
ing biological aggregationsin naturemust be to distinguish traffic rules-and quantitativelyassess the emergent prop-
biologically relevantfeaturesfrom nonadaptiveepiphenom- erties of the resulting schools. Ideally, both the inputs (i.e.,

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions
298 J. K. PARRISH ET AL

StScI, NASA Norbert Wu

OAR/ERL/NSSL NASA T. C. Schneirla

Figure 1. Three- and two-dimensionalexpressions of emergent structurein animate and inanimateaggre-

gations-whorl patterns.Left: tornadostructurein fish (top) and water vapor (bottom). Middle and right: whorl
and toroid in planets, fish, ants, and water vapor, clockwise from top left. Top left reprintedwith permission by
FPG; bottom left courtesy of National Severe Storm Laboratory;middle top and bottom courtesy of National
Aeronautic Space Administration;top right taken by NorbertWu, www.norbertwu.com? 1999; bottom right
taken by T. Schneirla and reprintedwith permission by W. H. Freeman& Company.

individuals' responses to neighbors) and model output specifically address hydrodynamic interactions between
(group-level characteristics)can be comparedto data from members of a school (e.g., Weihs, 1973).
real aggregations. Published modeling studies have for the most part ad-
A key purpose of modeling is to distinguish behavioral dressed a relatively restricted range of behavioral algo-
cause from organizational effect by studying the conse- rithms. Most analyses focus on variation in one of three
quences of various hypothetical social interaction rules. categories: behavioralmatching, positional preference, and
Most simulation models of animal aggregations in the lit- numericalpreference.
erature assign a set of forces that act on the speed and Behavioral matching (aka allelomimesis: Deneubourg
direction of each individual and are modulatedin response and Goss, 1989) occurs when the individual agents try to
to other individualsor the local environment.Typical force match their behavior with other nearby agents. Most often,
componentsinclude locomotory (e.g., biomechanicalforces behavioral matching is modeled by having each agent ex-
such as drag),aggregative(e.g., long-rangeattraction,short- plicitly match its orientationwith that of its nearby neigh-
range repulsion), arrayal(e.g., velocity matching), and ran- bors (e.g., a zone of parallel orientation:Aoki, 1982; Huth
dom (e.g., individual stochasticity; Griinbaumand Okubo, and Wissel, 1992; Dill et al., 1997; Table 1, Fig. 3). Less
1994). The detailed biomechanics of locomotory forces are often, fish match their speed, either to that of their compan-
usually not considered in fish schooling simulations. In- ions, or to some arbitrarilydecided value (e.g., Romey,
stead, most simulations simply associate behavioralmove- 1996; Vab0 and N0ttestad, 1997).
ment "decisions"with the movements that result. However, Positional preference refers to each fish having a pre-
there are exceptions, notably a few modeling studies that ferred position relative to one or more of its companions.

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions
 SELF-ORGANIZEDFISH SCHOOLS 299

Avoid f Herd
-C
Flashexpansion 7i~-;
_"F

_ZZ

Figure 2. Examples of coordinated movement and directed activity, both emergent properties of fish
schools, which are also commonly cited defense tactics against predatoryattack. Fig. 12.8. from Pitcher and
Parrish 1993. Reprintedwith permission from the authorand Kluwer Academic Publishers.

Usually, positional preference is formulatedas a preferred relatively small population (e.g., 8 to 20 fish; Table 1),
distance to one or more nearest neighbor(s). Variations on despite the fact that several studies point out that small
positional preferenceinclude assigned distance zones (e.g., numbersof fish may produce artificialresults (e.g., Romey,
repulsion, parallel orientation,attraction,searching) within 1996). Most models have operatedin two dimensions. Al-
which neighbors are treated equally (Huth and Wissel, though this may be justified on the basis of computational
1992; Stocker, 1999) or continuous distance weighting resources, generalizing from two to three dimensions may
(Warburtonand Lazarus,1991; Reuterand Breckling, 1994; not be straightforward.Many models include some stochas-
Romey, 1996). In some models, otherpositional parameters tic or chaotic elements;however, the degree of replicationin
influence responses, such as bearing angle to neighbors, or modeling studies is generally lower than ideal to character-
estimated collision time (Dill et al., 1997). The biological ize the frequency distributionsof possible outcomes.
underpinningis obvious: that individualgroup members do Most simulation studies also assume that all individuals
not collide, that groups do not dissolve, and that stragglers are identical. Romey (1996) has shown that the inclusion of
join. Taken together, behavioral matching and positional a single fish with different traffic rules will alter schooling
preference describe what a fish should do, e.g., move to- (measured as group speed and turning rate). Furthermore,
wards or away from neighbors,align with neighbors,search
the inclusion of multiple agents with one of two rule sets
for neighbors(Fig. 3; RPOA dependence,Table 1), and how
much consideration it should give any neighbor in the produces sorting (Romey, 1996), which may translateinto
emergentproperties(e.g., horizontalbands versus extended
perceptionfield (neighbor scaling rule, Table 1).
columns of ungulates, Gueron et al., 1996). The possibility
Numerical preference refers to the number of neighbors
to which a fish pays attention,which we generically refer to that variabilitywithin the group may not only affect emer-
as the rule size. Variationsinclude an a priori value (e.g., 4: gent propertiessuch as coordinatedmovement and spatial
Warburtonand Lazarus, 1991) or a conditional value (e.g., pattern,but may also itself be an emergentproperty,clearly
choose up to 4 in the nearest zone: Aoki, 1982; choose up deserves furtherconsiderationin modeling studies.
to 4, frontprioritized:Huthand Wissel, 1992). Many simply The parameterspace explored by each model is typically
have each fish average over all other fish (e.g., all visible: only a small subset of possible variations, which ideally
Reuter and Breckling, 1994; Vab0 and N0ttestad, 1997; all would include variations in (at least): initial position and
fish: Romey, 1996). velocity; the strength and type of stochastic components;
In general, modeling studies of schooling have been spatial distribution of repulsion, parallel orientation, and
limited in several importantrespects, which future studies attraction;and degree of variationbetween individuals in a
should aim to improve upon. Most models have used a group (Table 1). Because most studies are not consistent in

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions
300 J. K. PARRISH ET AL.

Table 1

Summaryoffish school simulationparameters and output variables. R/PO/A-sequentialzones of repulsion,parallel orientation (variably used), and
attraction where "zone" implies equal force within a proscribed area (see Figure 3). Direction matching implies the use of a zone of parallel
orientation.All other definitionsas in the text. Numbers in parentheses refer to footnotes.

Population R/PO/A Direction

Reference Size StartingOrientation Velocity Dependence Matching? Neighbor Scaling Rule Size

Aoki (1982, 1984) 8, 32 Random position & Random Zone (1, 2) Yes Position, weighted by Up to 4
orientationwithin front priority
bounded area
Warburton& Lazarus 2-9 Regular square lattice Not specified Lineardistance No Constant 1-8
(1991)
Huth & Wissel 8 Random position & Random Zone (1) Yes Constant;Single Up to 4, front
(1990, 1992, 1994) orientationin fixed choice (4) prioritized
area
Reuter & Breckling 10, 20, 30, Random position, Not clear Lineardistance Yes Distance-weighted All visible
(1994) 40,50 orientation& average (1/D; 1/D2)
speed within
bounded area
Romey (1996) 2-10 Random position & Constant Lineardistance No Constant All
orientationwithin
bounded area (5)
Vab0 & N0ttestad 900 Random position & Constant Discrete distance No Single choice (7) All visible
(1997) orientationwithin
bounded & fixed
areas
Stocker (1999) 12, 64 Random position & Constant Zone (1) Yes Constant Variable(8)
orientationwithin
bounded area

Random Component Aggregation Indices Aggregation Indices Aggregation Indices

Reference (distribution) Individual Group Population

Aoki (1982, 1984) Gamma (speed); Net IndividualDisplacement Compactness (1/SD of x, y

Normal (direction; 3) coordinates)
Warburton& Lazarus Circularnormal Inter-IndividualDistance Group Center Displacement;
(1991) (direction) Group Shape (max IID2)/area
Huth & Wissel Gamma (speed); Nearest Neighbor Distance Polarization;Expanse (rms
(1990, 1992, 1994) Normal (direction) (neighbors 1-3) distance to center)
Reuter & Breckling Normal (direction, Nearest Neighbor Distance Polarization;Expanse (mean
(1994) speed) distance to center)
Romey (1996) Uniform (direction;6) Group Speed; Group Turning
Rate
Vab0 & N0ttestad Discrete probabilitiesof Instability (maximum Number of Schools
(1997) choosing less than individual distance traveled)
optimal direction
Stocker (1999) Uniform (direction;9) Polarization;School Direction Number of Schools;
School Size

Note that random componentsselected from a normal distributionare centered on calculated means as per Figure 3.
(1) Distance-specific zones of repulsion, parallel orientation,attraction,and searching. A blind spot (30-60 deg.) behind.
(2) Beyond the zone of attraction,but within the visual angle (270 deg), attractionbased on linear distance.
(3) Outside the sensed area, direction was chosen from a uniform circulardistribution(i.e., randomwalk).
(4) Single Choice = choose only one neighbor with a weighted probabilityN1, 1/2N2, 1/4N3, 1/8N4.
(5) Individualscan sense entire area (-100 BL).
(6) Every 15 or 30 time steps.
(7) Highest (optimal) attractionvalue direction (of eight 45 deg increments)chosen.
(8) All neighbors within the parallel orientationzone considered equally. Only closest neighbor within attractionzone considered.
(9) Only applied to single fish with no neighbors (i.e., random walk searching).

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions
 SELF-ORGANIZEDFISH SCHOOLS 301

. (a) 1- (b) fore the entire population is always part of the same group,
0.5- / e.g., Aoki, 1982; Reuter and Breckling, 1994). Even when
this is not the case, most studies use initial conditions or
?io-0 / behaviors that predispose individuals to start and remain
-0.5 within a single large group. However, there are again nota-
-1 . , . ble exceptions, for example, the cellular automata models of
Vab0 and N0ttestad (1997) and St6cker (1999); see also
1 (c) (d) Flierl et al. (1999).
0.5;- I 1*Aa Inconsistenciesbetween modeling approachesand aggre-
__ ' < gation indices have led differentauthorsto draw seemingly
0 )- \
0.5-
-0.5
contradictoryconclusions about how individual behaviors
j/ v-^affect group-level characteristics.For example, Aoki (1982;
-1 0 . 1984) reports that removing the repulsion zone causes
school disintegration, whereas Huth and Wissel (1990,
I' (e) 1- 1994) assert it has little effect on schooling. Huth and
\(1992,
Wissel (1990, 1992, 1994) maintainthat only a low number
")u/ 0.5- of influential neighbors (< = 3) are needed, whereas War-
burtonand Lazarus(1991) state that increasingthe rule size
4/ ^ -- - is essential for school stability.Reuterand Breckling (1994)
and Romey (1996) found that adding randomor individual
{g)
(. 1- (h) movement had little effect on schooling, but Warburtonand
0.5
;
0.5- vYLazarus (1991) concluded that the incorporationof random
movements destroyedschooling. Because these studies lack
a common frameworkfor specifying behaviorsand summa-
-0.5 -0.5 - rizing spatial patterns, it is not possible to resolve whether
-1 .-1 . . these reflect true biological differences or simply compari-
0 1/3 D 23 D D 0 1/3 D 2/3 D sons between apples and oranges.
Figure 3. Caricaturesof the attraction-repulsionfunctions used by
various authorsto simulate schooling in fish (see also Table 1). Repulsion
A Synthetic Simulation Study
is always negative on the y-axis, whereas attractionis positive. Zones of no
repulsion or attraction(e.g., parallel orientation)are at zero. The x-axis is
We are currentlyworking on a simulation model of fish
distance from fish of interest,displayed in arbitraryunits. (a) Aoki (1982).
(b) Warburtonand Lazarus (1991). (c) Huth and Wissel (1990, 1992, schooling that explores much of the parameterspace cov-
1994). (d) Curves showing the relative strengthsof repulsion(left), parallel ered in previous models. Our model operates in three di-
orientation(center), and attraction(right) in Reuter and Breckling (1994). mensions, and does not (as yet) include any rules that, a
Note that total social force always sums to 1.0 (e) Romey (1996). (f) The
priori, programemergentproperty(i.e., no behaviormatch-
attractionstrengthused by Vab0 and N0ttestad (1997). (g) St6cker (1999).
(h) Our simulation.
ing or zone of parallel orientation).Following Grunbaum
and Okubo (1994) and others, we model the movement
decisions of each fish as a sum of locomotory, social, and
these aspects (among others), direct comparisons of their random forces. Model fish are based on our quantitative
results are not possible. observationsof schooling giant danios (Danio aequipinatus;
Finally, although most authorsquantifiedschool charac- unpub. data), with a mass of 1.7 g, a length of 5.3 cm, a
teristics in their simulations with one or more index of maximum speed of 11.2 body lengths [BL]/s, and a maxi-
aggregation,there is little or no consensus on which indices mum acceleration of 11.2 BL/s2). Locomotory force con-
are most biologically relevant.Consequently,differentstud- sists of drag, specified as a drag coefficient (CD)of 0, 0.01,
ies tend to cite different statistics, again making compari- or 0.02 (based on a range of estimateddrag coefficients for
sons difficult. For example, it is difficult to determine streamlinedbodies = 0.01-0.001; Videler, 1993). Details
whether Romey's (1996) model provides the same amount of the methodology are presentedelsewhere (Viscido et al.,
of polarization as Huth and Wissel's (1992), because the unpub. data).
former did not report that particularemergent property.A We have thus far used this model to study a range of
related but more subtle problem is that no single paper has population from 2 to 128 fish, but reporthere only results
simultaneously presented indices at the individual, group, for 128 fish. We examined variationin three aspects of the
and population levels (Table 1). In part this is because the social force: the attraction/repulsionfunction, the neighbor
distinctionbetween group and populationis meaningless in scaling rule, and the rule size. Attraction/repulsionwas
some models (i.e., all individualsalways interactand there- modeled as a linear, piecewise linear, or convex (sensu

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions
302 J. K. PARRISH ET AL.

Table 2

Aggregation indices at the individual, group, or population level used to evaluate traffic rules in our simulation

Index Computation Level

Curvature Mean turningangle per cm traveled Individual

FractalDimension How "dimensional"the path is Individual
NGDR Net displacementdivided by gross displacement Individual
NND Mean linear distance to nearest neighbor Individual
Group Size Number of fish within 5 body lengths of each other Group
Expanse Mean quadraticdistance to center for each group Group
Polarization Mean angle deviation between individual and group heading Group
Groups Number of groups in the universe, including stragglers Population
Stragglers Number of fish more than 5 body lengths from any other fish Population
Collisions Number of collisions per fish per minute (reportedas total collisions) Population
Clark and Evans (1954) Index Observed NND vs. expected NND for a randomly spaced population Population

Warburtonand Lazarus, 1991) function with F = 0 cen- rule size of 16. All numbers in text are means, and all
tered at 1.9 BL (Fig. 3h). We based this preferrednearest statistics are one-way ANOVAs.
neighbor distance on schooling experiments with real fish In general, over the range of parametervalues we used,
from a previous study (Parrishand Turchin,1997). We used asocial forces (drag,randomness)had a less drasticeffect on
a perception distance of 180 BL (i.e., so individuals could fish schooling behaviorthan did the various functions com-
detect neighbors to which they choose not to respond). A posing the social force (Fig. 4). Within the range of random
neighborscaling rule weighted the attraction/repulsionforce force examined,there was no significanteffect on any index
by the distance to each influentialneighbor.Weighting was of aggregation (Fig. 4). The addition of drag slowed fish
constant (i.e., equal weighting), negative linear, or negative down and increased the tortuousity of their paths, with a
sigmoid. Finally, the number of influentialneighbors (rule resulting increase in group size (without drag-46 fish/
size) was set at 4, 8, 16, or 24 fish. group, with drag-104 fish/group) and cohesion (without
Random force was chosen for each timestep from a drag-1.5BL, with drag-1.1BL). These larger, slower
normal distributionwith mean F = 0 and standarddevia- groups milled rather than moving forward (net to gross
tion equivalentto 1/3, 1/6, and 1/18 of the maximum social displacementratio [NGDR] droppedfrom 0.5 to 0.1), which
force an individual fish could experience from anotherfish. decreasedpolarization.However, once added, there was no
The amountof randomforce experiencedby any individual differencebetween the low and high dragcondition (Fig. 4).
was a function of the distribution of nearby neighbors. Social forces had a significanteffect on most measuresof
Individualswithin a school experienced relatively little so- schooling. Fish with convex attraction-repulsion(A/R) re-
cial force because most neighbors are at or near the equi- sponses stayed closer togetherand had more tortuouspaths
librium (i.e., F = 0), thus random forces can be propor- than fish with piecewise linear A/R responses. The result
tionately large. Individuals divorced from the group but was an increase in collisions (convex-720 collisions;
close enough to sense it experiencedmaximumsocial force piecewise linear-270 collisions; Fig. 4). Polarizationwas
(i.e., attraction),and proportionatelylittle randomforce. In unaffected;it was equally low (83 deg.) for both functions.
our simulations,median randomforce as a function of total Numericalpreferencehadstrongeffects,bothin termsof the
force ranged from 8%-18% depending on rule size. We numberof influentialneighborsand how that influencewas
computed fish movement as accelerationresulting from the scaled.Changingthe neighborscalingrulefromequalweight-
sum of random and social forces, minus drag. ing to a negative sigmoid function effectively increasedthe
Each parameter combination in our simulations were influence of close neighbors. Fish moved faster and in
replicated 15 times, for 1800 time steps (of 1 s) each. We straighterpaths. Schools became smaller (sigmoid-33 fish/
examined our model output with a suite of 11 aggregation group;constant-104 fish/group)and slightly more polarized
indices thatencompass individual-to population-levelchar- (sigmoid-70 deg.; constant-83 deg.), but also more diffuse
acteristics (Table 2). Statistics were based on time averages (sigmoid-1.9BL; constant-l.lBL). As a result, collisions
taken over the second half of each simulation (900 time decreased(sigmoid-5 collisions; constant-270 collisions).
steps), to avoid undue influence from initial conditions. We Using a smallerrulesize (4 fish)producedequivalentresultsto
report here the results of systematic comparison within a negative sigmoid neighborscaling. Thus a highernumberof
single rule (e.g., dragoff versus high drag)where the default influentialneighborsincreasesschool size at the expense of
conditions were medium random force, no drag, piecewise internalstructure(measuredas polarization).Paying attention
linear attraction/repulsion,constant neighbor scaling, and a to more neighborsleads to large groups because it causes a

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions
 SELF-ORGANIZEDFISH SCHOOLS 303

Index RandomForce Drag A/R function Neighbor Rule Size ally identical (Fig. 5b), indicatingthat at this scale, spacing
Low High Convex Scaling 4
(High) (Low) (Piecewise Logistic (16) is quite regular.However, when rule size is very large (24
linear) (Constant) influential neighbors), this relationship breaks down (Fig.
Curvature O t t f 5b), perhapsindicating an upper limit for rule size.
Although our results are preliminary,they highlight some
FractalD 0> > > interestingpatterns.First, school cohesion and coordinated
movement can be partially achieved by positional and nu-
NGDR K> K> I merical preferences,without the need for behavioralmatch-
ing. This shows that schools might indeed arise by large

If
NND
0 0 I numbers of simple, overlapping interactions (Warburton
and Lazarus, 1991; Romey, 1996). Our preliminaryresults
Grp Size C> 0 l also indicate that the shapes of the functions used to char-
acterize both positional and numerical preferences are
Expanse K K> l t
clearly very important(Fig. 4), and worthy of careful in-
vestigationin the future.Finally, our resultsdemonstratethe
Polar o fo f need to consider many different quantitative indices of
aggregationwhen assessing model results (Fig. 4).
groups O ? t t
Emergent Evolution
#strag O ? ? ? f
A central issue in modeling studies of aggregation is
#crashesO 1I distinguishingbetween group- and population-levelcharac-
teristics that are direct, inevitable consequences of the as-
IoA O> ?
fl sumptions built into the model, and those that are truly
"emergentproperties."In an evolutionarycontext, the im-
Figure 4. Results of one-way ANOVAs systematically comparing
parametervalues within a given traffic rule listed as column headings.
Withineach trafficrule, arrowsindicatea significantdifferencein the index
0'
between the listed condition (e.g., rule size, "4" nearestneighbors) to that ORS = 4
of the default set in parentheses(one-way ANOVA, n = 15, P < 0.001). 10 - *RS = 8 (A)
ORS = 16
White diamonds indicate no significantdifference. Rows are 11 indices of o 20-
G0 ARS = 24
aggregation. Individuallevel: path curvature,fractal dimension (higher is
30
more tortuous), net to gross displacement ratio (NGDR; higher is
straighter),nearest neighbor distance (NND). Group level: group size (in .2 40 d""
fish), mean distance to group center (or expanse, in body lengths), polar- - 50 Us
A
ization (lower is more aligned). Population level: number of groups,
? 60
numberof collisions, numberof stragglers,and the Clarkand Evans (1954) aS
A
index. See Table 2 for details on how each index was computed. 70

An
0 2 4 6 8

Group Speed (BL/s)

"network of overlapping influences" (Warburton and Lazarus,
1991): each group member "pulls on" a different set of com- _ 8
z (B)
panions, causing groups to remain cohesive. Over the range of ,L 7 *
z
rule size (RS) examined,
Q 6-
Z 5
Group Size = 3 RS - 7 (df = 59, r2 = 0.85). 24
..~RS=
-
4 ^NAN\^
~~~vw
Several other higher order properties emerged during our r 3-

modeling. There was a clear linear relationship between 2 RS= 16

group speed and polarization, across all rule sizes. Faster S

/
I--RS= 8
- -- - - - -- .- - -- -
c mmcmmlgw
groups are more polarized (Fig. 5a). We note that there is no
30 35 40 45 50 55 FS=4
preferred speed, so that this is not an inevitable result of a Time (seconds)
polarized group, but probably reflects the fact that slower
groups are composed of individuals with more tortuous Figure 5. Higher-orderindex comparisons. (A) Group speed versus
polarizationfor rule sizes 4, 8, 16, and 24. (B) Mean variance in the six
paths. We also have begun to examine substructure within nearest neighbordistances for individuals in the center of the school, over
groups. At the level of an individual's immediate neighbors four rule sizes. Individualswere considered"central"if they were closer to
(here defined as N1-6), nearest neighbor distance is virtu- the group center than the median distance to center for the whole group.

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions
304 J. K. PARRISH ET AL

portanceof this distinction is that emergent propertiesmay Simulation studies like these hinge directly on two ques-
be closely tied to indirect behavioral adaptations, which tions (1) what is optimal, and (2) what is the relationship
benefit individuals by benefiting the group. To satisfy the between individual and group optimality? Such questions
definition of emergent properties, the larger system (the are certain not to have simple or universal answers. It is
group) must possess them, while its components (the indi- unlikely that we will soon attaincomplete, realistic models
vidual members) do not (Clark et al., 1997). To a large of an aggregatingorganism's life history that would allow
extent, designating a group characteristicas "emergent"or us to truly represent"fitness"in a populationwith evolving
not is a matter of degree. For instance, models in which behaviors. However, we are optimistic about the role of
individualsactively align their directionsto those of neigh- schooling simulations in addressing two central issues
bors typically produce polarized groups (Huth and Wissel, raised in this paper.First, we believe model studies will be
1992, 1994; Reuter and Breckling, 1994; Stocker, 1999), successful in elucidating the mechanistic relationshipsbe-
but this is probably too direct an outcome of the assump- tween individual-level behavior and the group-level spatial
tions to be considered an emergent property. Dill et al. patternsthey produce. Second, we expect that the combi-
(1997) pose a model in which each fish estimatesits time to nation of explicit observationsof schooling fish, improved
impact with other fish. In such a system, agents implicitly simulationsof the observed behaviors,and expandeduse of
consider their neighbors' orientation,velocity, and position, statistical indices of aggregationon real and model aggre-
ratherthan explicitly doing so. Collision avoidance would gations will prove sufficient to identify the true behavioral
thereforeseem to be a direct outcome and not an emergent algorithmsused by fish.
propertyof this model. However, what about the converse:
is polarizationan emergent propertyof collision-avoidance
Acknowledgments
behavior?
From a biological perspective, an operationaldistinction This paper is the result of fruitful interactionswith the
might be to consider whether the group behavior benefits participantsof the CASSLS workshop (The Limits to Self
the members because of their membership.For example, Organization in Biological Systems) and the International
Gruinbaum(1998) used model results to suggest that when Ethological Conference symposium (The Many Facets of
individuals simultaneously display gradient-climbingand Gregariousness),in particularJean-LouisDeneubourg,Guy
alignment behaviors in a noisy environment, groups to Theraulaz,CharlotteHemelrijk,and LaurentDagorn. Kate
which they belong more accurately climb environmental Litle and Jennifer Weitzman provided invaluable last-
gradients,whereas loners can not. If displaying both behav- minute help. The original manuscriptwas greatly improved
iors is costly to an individual,but it nonetheless benefits by by two anonymousreviewers. This researchwas supported
being part of a group in a better environment,this would in part by a National Science Foundationgrant to JKP and
qualify as emergent under this criterion. DG, CCR-9980058.
Given the relatively underdevelopedstate of schooling
behavior models and observations that unify individual, Literature Cited
group, and populationcharacteristics,it is unavoidablethat
studies of the evolutionary "hows and whys" of schooling Aoki, I. 1982. A simulation study on the schooling mechanism in fish.
Bull. Jpn. Soc. Sci. Fish. 48: 1081-1088.
remain somewhat speculative and oversimplified.Nonethe-
Aoki, I. 1984. Internaldynamics of fish schools in relation to inter-fish
less, several studies have attemptedto incorporateevolu- distance. Bull. Jpn. Soc. Sci. Fish. 50: 751-758.
tionary ideas into schooling models by allowing iterative Aoki, I., T. Inagaki, and L. van Long. 1986. Measurementsof the
improvementof parameters,using an objective function as three-dimensionalstructureof free-swimmingpelagic fish schools in a
an analog to "fitness"in the evolutionary sense (Huse and naturalenvironment.Bull. Jpn. Soc. Sci. Fish. 52: 69-77.
Bonabeau, E., G. Sylvain, D. Snyers, P. Kuntz, and G. Theraulaz. 2000.
Gjoesaete, 1999; Bonabeauet al., 2000). For instance,Hira- Three-dimensionalarchitecturesgrown by simple 'stigmergic' agents.
matsu et al. (2000) link a standardsimulation of schooling Biosystems 56: 13-32.
(a la Huth and Wissel, 1992) to a genetic algorithm that Camazine, S., J.-L. Deneubourg, N. R. Franks, J. Sneyd, G. Theraulaz,
selects the width of the parallelorientationzone, the size of and E. Bonabeau. 2001. Self-Organizationin Biological Systems.
the blind angle, the numberof influentialneighbors,and the PrincetonUniversity Press, Princeton,NJ.
Clark, P. J., and F. C. Evans. 1954. Distance to nearest neighbor as a
neighbor scaling rule (front, side, or distance priority,ran- measureof spatial relationshipsin populations.Ecology 35: 445-453.
dom choice) to maximize fitness. In this case, fitness is Clark, M., K. Irwig, and W. Wobke. 1997. Emergent properties of
modeled as the minimizationof four outputvariables:near- teams of agents in the Tileworld. Lect. Notes Comput. Sci. 1209:
est neighbor distance, polarization, expanse, and fractal 164-176.
dimension. As a result, parallel orientation zone is maxi- Deneubourg, J. L., and S. Goss. 1989. Collective patternsand decision-
making. Ethol. Ecol. Evol. 1: 295-311.
mized and blind angle is minimized. A larger question is
Dill, L. M., C. S. Holling, and L. H. Palmer. 1997. Predicting the
whether this definition of most fit-that is, least vari- three-dimensional structure of animal aggregations from functional
able-is correct? considerations:the role of information.Pp. 207-224 in Animal Groups

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions
 SELF-ORGANIZEDFISH SCHOOLS 305

in ThreeDimensions, J. K. Parrishand W. M. Hamner,eds. Cambridge Parrish, J. K. 1993. Comparisonof the huntingbehaviorof four piscine
University Press, New York. predatorsattackingschooling prey. Ethology 95: 233-246.
Flierl, G., D. Griinbaum, S. Levin, and D. Olson. 1999. From individ- Parrish, J. K., and L. Edelstein-Keshet. 1999. Complexity,pattern,and
uals to aggregations:the interplay between behavior and physics. J. evolutionarytrade-offs in animal aggregation.Science 284: 99-101.
Theor. Biol. 196: 397-454. Parrish, J. K., and L. Edelstein-Keshet. 2000. Reply to Danchin and
Griinbaum, D. 1998. Schooling as a strategy for taxis in a noisy envi- Wagner. Science 287: 804-805.
ronment.Evol. Ecol. 12: 503-522. Parrish, J. K., and P. Turchin. 1997. Individualdecisions, trafficrules,
Griinbaum, D., and A. Okubo. 1994. Modelling social animal aggre- and emergentpatternin schooling fish. Pp. 126-142 in Animal Groups
gations. Pp. 296-325 in Frontiers in Theoretical Biology: Lecture in ThreeDimensions, J. K. Parrishand W. M. Hamner,eds. Cambridge
Notes in Biomathematics,Vol. 100, S. A. Levin, ed. Springer-Verlag, University Press, New York.
New York. Partridge, B. L., and T. J. Pitcher. 1980. The sensory basis of fish
Gueron, S., S. A. Levin, and D. I. Rubenstein. 1996. The dynamics of schools: relative roles of lateralline and vision. J. Comp.Physiol. 135:
herds: from individuals to aggregations.J. Theor. Biol. 182: 85-98. 315-325.
Hillier, B., and J. Hanson. 1990. The Social Logic of Space. Cambridge Partridge, B. L., J. Johansson, and J. Kalisk. 1983. Structure of
University Press, New York. schools of giant bluefin tuna in Cape Cod Bay. Environ.Biol. Fishes 9:
Hiramatsu, K., S. Shikasho, and K. Mori. 2000. Mathematicalmod- 253-262.
eling of fish schooling of Japanese medaka using basic behavioral
Pitcher, T. J., and J. K. Parrish. 1993. Functionsof shoaling behavior
patterns.J. Fac. Agric. Kyushu Univ. 45: 237-253. in teleosts. Pp. 363-439 in Behaviour of Teleost Fishes, 2nd ed.
Huse, G., and H. Gjoesaete. 1999. A neural network approach for
Chapmanand Hall, London.
predicting stock abundance of the Barents Sea capelin. Sarsia 84:
Reuter, H., and B. Breckling. 1994. Self-organizationof fish schools:
457-464.
an object-orientedmodel. Ecol. Model. 75: 147-159.
Huth, A., and C. Wissel. 1990. The movement of fish schools: a sim-
ulation model. Pp. 578-590 in Biological Motion: Lecture Notes in Romey, W. L. 1996. Individual differences make a difference in the
Biomathematics, Vol. 89, W. Alt and G. Hoffman, eds. Springer, trajectoriesof simulated schools. Ecol. Model. 92: 65-77.
Berlin. Schilt, C. R., and K. S. Norris. 1997. Perspectives on sensory integra-
tion systems: problems,opportunities,and predictions.Pp. 225-244 in
Huth, A., and C. Wissel. 1992. The simulationof the movement of fish
schools. J. Theor. Biol. 156: 365-385. Animal Groupsin ThreeDimensions,J. K. ParrishandW. M. Hammer,
Huth, A., and C. Wissel. 1994. The simulation of fish schools in eds. CambridgeUniversity Press, Cambridge,UK.
comparisonwith experimentaldata. Ecol. Model. 75: 135-145. Shaw, E. 1978. Schooling fishes. Am. Sci. 66: 166-175.
Latan6, B., J. Liu, A. Nowak, M. Bonavento, and L. Zheng. 1995. Stocker, S. 1999. Models for tuna school formation.Math. Biosci. 156:
Distance matters:physical space and social influence.Pers. Soc. Psych. 157-190.
Bull. 21: 795-805. Vab0, R., and L. N0ttestad. 1997. An individual-basedmodel of fish
Norris, K. S., and C. R. Schilt. 1988. Cooperative societies in three- school reactions: predicting antipredatorbehaviour as observed in
dimensional space: on the origins of aggregations,flocks, and schools, nature.Fish. Oceanogr. 6: 155-171.
with special reference to dolphins and fish. Ethol. Sociobiol. 9: 149- Videler, J. J. 1993. Fish Swimming.Chapman& Hall, London. 260p.
179. Warburton, K., and J. Lazarus. 1991. Tendency-distancemodels of
Parrish, J. K. 1989. Re-examiningthe selfish herd:are centralfish safer? social cohesion in animal groups. J. Theor. Biol. 150: 473-488.
Anim. Behav. 38: 1048-1053. Weihs, D. 1973. Hydromechanicsof fish schooling. Nature 241: 290-
Parrish, J. K. 1992. Do predators 'shape' fish schools: interactions 291.
between predatorsand their schooling prey. Neth. J. Zool. 42: 358- Wilson, E. 0. 1975. Sociobiology: The New Synthesis.HarvardUniver-
370. sity Press, Cambridge.672p.

This content downloaded from 193.142.30.55 on Sat, 28 Jun 2014 08:34:25 AM

All use subject to JSTOR Terms and Conditions