Emotional Gestalts: Appraisal, Change, and the Dynamics of Affect

Paul ThagardPhilosophy DepartmentUniversity of Waterloo

Josef NerbPsychology DepartmentUniversity of Freiburg

This article interprets emotional change as a transition in a complex dynamical sys-tem. We argue that the appropriate kind of dynamical system is one that extends recentwork on how neural networks can perform parallel constraint satisfaction. Parallelprocesses that integrate both cognitive and affective constraints can give rise to statesthat we call emotional gestalts, and transitions can be understood as emotional ges-talt shifts. We describe computational models that simulate such phenomena in waysthat show how dynamical and gestalt metaphors can be given a concrete realization.

One Tuesday morning, Professor Gordon Atwoodstrode energetically into his department office at Yu-kon University. It was a great day—the sun was shin-ing, his family had been good natured at breakfast, andhe was finished teaching for the semester. He cheerilygreeted the department secretary and went to his mail-box, eagerly noticing the long-awaited letter from theJournal of Cognitive Chemistry. After opening the let-ter, Gordon’s heart sank as he read the dreaded words:“We regret to inform you that, based on the reviewers’reports, we are unable to accept your submission.” Ashe scanned the reviews, sadness turned to anger whenGordon realized that one of the reviewers had totallyfailed to understand his paper, while the other had re-jected it because it did not sufficiently cite the re-viewer’s own work. Stomping out of the departmentoffice, Gordon chided the secretary for running out ofprinter paper yet again.

After an unproductive morning largely spent surf-ing the Web, Gordon met his wife for lunch. She re-minded him that he had recently had 2 articles acceptedfor publication, and the rejected article could always bepublished elsewhere. They chatted about their daugh-ter’s recent triumph in a ballet school performance, andmade plans to see a musical that weekend. The Thaicurry was delicious, and the coffee was strong. Gordonreturned to work and happily updated his article forsubmission to a different journal.

Everyone has had days like this, with transitions be-tween different moods and emotions. A psychologicaltheory of affect must explain both how different emo-tional states arise and how one state can be replaced byanother one that is qualitatively very different. Affect isa natural subject for a dynamical theory that empha-sizes the flow of thought and the complex interactionsof emotion and cognition. Our aim in this article is todevelop such a theory of the emergence and alterationof emotional states.

We proceed first by interpreting emotional changeas a transition in a complex dynamical system. Thismetaphorical interpretation, however, is limited in itsexplanatory power without concrete specification ofthe structures and mechanisms that can give rise toemotions and emotional change. We argue that the ap-propriate kind of dynamical system is one that extendsrecent work on how neural networks can perform par-allel constraint satisfaction. Parallel processes that in-tegrate both cognitive and affective constraints cangive rise to states that we call emotional gestalts, andtransitions such as those experienced by Gordon can beunderstood as emotional gestalt shifts. Finally, we de-scribe computational models that simulate such phe-nomena in ways that show how dynamical and gestaltmetaphors can be given a concrete realization.

Emotion as a Dynamical System

At its simplest, psychological theory postulatescausal relations between mental properties and behav-ior, for example, that there is a personality trait ofextraversion that leads people who have it to talk fre-quently with other people. Since the rise of cognitivescience in the 1950s, psychological theories have in-

Personality and Social Psychology Review2002, Vol. 6, No. 4, 274–282

Copyright © 2002 byLawrence Erlbaum Associates, Inc.

274

This research was supported by a Natural Sciences and Engi-neering Research Council of Canada grant to Paul Thagard, and by afellowship from the Alexander von Humboldt-Stiftung to Josef Nerb.We are grateful to Stephen Read, Robin Vallacher, Eliot Smith, and ananonymous referee for helpful comments on an earlier draft.

Requests for reprints should be sent to Paul Thagard, PhilosophyDepartment, University of Waterloo, Waterloo, ON, Canada, N2L3G1. E-mail: pthagard@uwaterloo.ca

creasingly postulated representational structures andcomputational processes that operate on the structuresto produce behavior. More recently, some psychologistsand philosophers have proposed that psychological the-ories should be analogous to theories of complex dy-namical systems that have been increasingly popular inphysics, biology, and other sciences (see, for example,Port & van Gelder, 1995; Thelen & Smith, 1994).

Thagard (1996, ch. 11) described how dynamicalsystems theory can be applied to psychological phe-nomenabymeansof thefollowingexplanationschema:

Explanation Target: Why do people have stablebut unpredictable patterns of behavior?Explanatory Pattern

Human thought is describable by a set ofvariables.

These variables are governed by a set of non-linear equations.

These equations establish a state space thathas attractors.

The system described by the equations ischaotic.

The existence of the attractors explains stablepatterns of behavior.

Multiple attractors explain abrupt phase tran-sitions.

The chaotic nature of the system explainswhy behavior is unpredictable.

It is easy to apply this explanation schema to emo-tions. We want to be able to explain both why peoplehave ongoing emotions and moods, and also how theycan sometimes make dramatic transitions to differentemotional states. Hypothetically, we might identify aset of variables that describe environmental, bodily,and mental factors. Equations that describe the causalrelations among those variables would clearly be non-linear, in that they would require specification of com-plex feedback relations between the different factors.The system described by the equations would undoubt-edly be chaotic, in the sense that small changes to thevalue of some variables could lead to very largechanges in the overall system: It only took one event todramatically change Gordon’s mood. On the otherhand, the emotional dynamic system does have somestability, as people maintain a cheerful or terrible moodover long periods. This stability exists because the sys-tem has a tendency to evolve into a small number ofgeneral states called attractors, and the shift from onemood to another can be described as the shift from oneattractor to another.

Compare the kinds of perceptual transitions thatwere identified by the gestalt psychologists. When yousee a Necker cube or a duck-rabbit, you see more thanjust the lines that make up the figure. The cube flipsback and forth as you see it in different gestalts, and the

duck-rabbit appears to you as a duck or as a rabbit, butnot both. Attending to different aspects of the drawingproduces a gestalt shift in which you move from oneconfiguration to the other. In the language of dynami-cal systems theory, the perceptual system has two at-tractor states, and the gestalt shift involves a phasetransition from one attractor to the other. Analogously,we might think of an emotional state as a gestalt thatemerges from a complex of interacting environmental,bodily, and cognitive variables, and think of emotionalchange as a kind of gestalt shift.

So much for metaphor. If we want a scien-tific–rather than a literary–explanation of emotionaland perceptual phenomena, we need to flesh out thedynamical systems explanation schema by specifyingthe variables and the equations that relate them. We canthen use computer models to simulate the behavior ofthe mathematical system to determine how well itmodels the complex behaviors of the psychologicalsystem under investigation. In particular, we want toknow whether the dynamical theory can explain bothstability and change in the psychological system, forexample both gestalts and gestalt shifts. We will nowdescribe how theories of thinking as parallel constraintsatisfaction can provide the desired explanations.These theories are implemented by connectionist (arti-ficial neural network) models with variables and equa-tions that define the behavior of dynamical systems.

Thinking as Parallel ConstraintSatisfaction

Thinking has traditionally been conceived as a se-rial process in which rules are applied to mental repre-sentations to generate new representations. In contrast,many kinds of thinking can be usefully understood assimultaneously involving numerous representationsthat constrain each other, with processes that operate inparallel to maximize constraint satisfaction. For exam-ple, Kunda and Thagard (1996) explained how peopleform impressions of other people from stereotypes byproposing that representations of different stereotypi-cal features, traits, and behaviors interact to produce anoverall impression or gestalt. Your overall impressionof someone you encounter as a male, Canadian, irrita-ble professor will depend on the simultaneous interac-tion of your representations for each of these concepts.Read, Vanman, and Miller (1997) have reviewed howrecent parallel constraint satisfaction theories have il-luminated the relevance of traditional gestalt principlesto social psychology. Many cognitive phenomena suchas analogy and categorization can be understood interms of parallel constraint satisfaction (see, for exam-ple, Holyoak & Spellman, 1993; Thagard, 2000).

What takes this approach beyond the metaphoricalis the availability of computational models that show

275

EMOTIONAL GESTALTS

how parallel constraint satisfaction can be performedby artificial neural networks. Representations such asconcepts and propositions can be modeled by units,which are artificial neurons with variables that indicatethe degree of activation (plausibility or degree of ac-ceptance) of each unit. The constraints between repre-sentations can be represented by excitatory and inhibi-tory links between units; for example, the positiveconstraint between the concepts professor and intelli-gent can be captured by an excitatory link between thecorresponding units, while the negative constraint be-tween professor and rich can be captured by an inhibi-tory link. Parallel constraint satisfaction is performedby means of equations that specify how the activationof each unit is updated as a function of the activationsof the units to which it is linked and the strength (posi-tive or negative) of those links. Updating usually leadssuch systems into stable states in which activations ofall units cease to change; this is called settling. How-ever, the systems are chaotic in that slight changes toinputs can lead the system to settle into a differentstate. For example, it is easy to model the Necker cubeby a neural network in which each unit stands for a hy-pothesis about which feature of the cube is front orback (Rumelhart & McClelland, 1986, vol. 2, p. 10).Slight changes to input about a particular feature beingfront or back can produce a flip in the overall state ofthe network. Thus artificial neural networks that imple-ment parallel constraint satisfaction can naturallymodel gestalts and gestalt shifts. We will now describehow parallel constraint satisfaction and connectionistmodeling can be extended to emotional thinking.

Emotional Gestalts: Theory

Extending parallel constraint satisfaction to emo-tion requires representations and mechanisms that gobeyond those required for purely cognitive coherence.In addition to representations of propositions, con-cepts, goals, and actions, we need representations ofemotional states such as happiness, sadness, surprise,and anger. Moreover, the cognitive representationsneed to be associated with emotional states, most gen-erally with positive and negative evaluations or va-lences (Bower, 1981; Lewin, 1951). For example,Gordon Atwood’s belief that his paper has been re-jected has negative valence and is associated with sad-ness, whereas his belief that his daughter is an excel-lent dancer has positive valence and is associated withpride. Like beliefs, concepts can also have positive va-lences, for example ones representing success and sun-shine, while other concepts have negative valences, forexample ones representing death and disease.

In purely cognitive coherence, representations areconnected by positive and negative constraints repre-senting the extent to which they fit with each other.

Similarly, representations can be connected by affec-tive constraints. Some of these are intrinsic to theperson that has the representations: most children, forexample, attach an intrinsic positive valence to candy.In other cases the valence is acquired by associations,as when a child acquires a positive valence for grocerystore because grocery stores sell candy.

In purely cognitive coherence models, representa-tions are accepted or rejected on the basis of whetherdoing so helps to satisfy the most constraints, where apositive constraint is satisfied if the representations itconnects are both accepted or both rejected. A judg-ment of emotional coherence requires not only infer-ences about the acceptance and rejection of representa-tions, but also about their valence. Gordon is forced toaccept the conclusion that his paper will not appear inthe journal, and he attaches negative valence to thisproposition. Emotional coherence requires not only theholistic process of determining how to best satisfy allthe cognitive constraints, but also the simultaneous as-sessment of valences for all relevant representations.

The result can be an emotional gestalt consisting ofa cognitively and emotionally coherent group of repre-sentations with degrees of acceptance and valencesthat maximally satisfy the constraints that connectthem. Of course, circumstances may prevent theachievement of coherence, for example, when there issupport for inconsistent beliefs or when competing ac-tions are associated with conflicting valences. But nor-mally the process of parallel constraint satisfaction willlead a person to acquire a set of stable acceptances ofpropositions and concepts that apply to the current sit-uation, as well as a set of valences that indicate theemotional value of those representations. Particularemotions such as happiness and sadness can thenemerge from an assessment of the extent to which thecurrent situation satisfies a person’s ongoing goals.

Emotional gestalt shifts occur when changes in rep-resentations and their valences generate a new array ofacceptances and valences that maximize constraint sat-isfaction differently from before. When Gordon readshis letter from the journal, he has to shift the cognitivestatus of the proposition that his article will be pub-lished in it from accepted to rejected. Through parallelconstraint satisfaction, this shift may alter the accep-tance status of other propositions, for example onesrepresenting whether he will get promoted or receive agood raise this year. In addition, there may be a shift invalences attached to his representations of the article,the journal, or even his whole career.

The theory of emotional coherence is highly com-patible with the appraisal theory of emotion, accord-ing to which emotions are elicited by evaluations ofevents and situations (Scherer, Schorr, & Johnstone,2001). Appraisal theorists have explained in great de-tail how different cognitive evaluations produce manydifferent emotions. But they have been vague about

276

THAGARD & NERB

the specific mechanisms that generate evaluations andelicit emotions. We propose that appraisal is a pro-cess of parallel constraint satisfaction involving rep-resentations that have valences as well as degrees ofacceptance, and we will describe in the next sectionneural network models that are capable of performingsome kinds of appraisal. According to appraisal the-ory, changes in emotion such as those that occur inpsychotherapy are due to changes in how individualsevaluate their situations; the theory of emotional co-herence gives a more specific account of how cogni-tive therapy works, by the therapist introducing newevidence and reforming coherence relations in waysthat change emotional appraisals (Thagard, 2000, p.209). Let us now consider more concretely how pro-cesses of parallel constraint satisfaction can produceemotional gestalts and affective change.

Emotional Gestalts: Models

Computational models are an indispensable tool fordeveloping psychological theories and for evaluatingtheir explanatory and predictive power. The view ofthinking as parallel constraint satisfaction did not pre-cede computational modeling, but rather arose becauseof the development of artificial neural network(connectionist) models (Rumelhart & McClelland,1986). It is useful conceptually to distinguish amongthe following: (1) a psychological theory of the struc-tures and processes that produce a kind of thinking, (2)a computational model that rigorously spells out thestructures and processes in terms of data structures andalgorithms, and (3) a running computer program thatuses a particular programming language to implementthe model and serves to test the power and empiricalrelevance of the model and theory. In practice, how-ever, the development of computer programs can be amajor part of the creation of new models and theories.

Computer simulations are crucial for theories ofcomplex dynamical systems. Metaphorical use of dy-namical system terminology (nonlinear systems,chaos, attractors, self-organization, emergence, etc.)can be useful for reconceptualizing complex phenom-ena, but a deep understanding of the phenomena re-quires mathematical specification of the key variablesthat produce them and of the relations among them.Once this specification is available, it is rarely possibleto infer the behavior of the system characterized bymathematical deduction or hand simulations, so com-puter simulations are needed for determining the im-plications of the mathematical model. As Vallacher,Read, and Nowak (this issue) emphasize, dynamicalresearch in psychology needs computer modeling as acomplement to empirical research.

Fortunately, implemented artificial neural networkmodels of emotional cognition are already available.

We will describe how two models, HOTCO andITERA, provide partial realizations of emotional ge-stalts, and indicate some directions for future work.Computer simulation of emotional or “hot” cognitionoriginated with the work of Abelson (1963) on ratio-nalization, but current models use artificial neural net-works to integrate cognition and emotion.

HOTCO

The model HOTCO (for “hot coherence”) incorpo-rates previous computational models of analogy, deci-sion making, hypothesis evaluation, and impressionformation (Holyoak & Thagard, 1989; Kunda &Thagard, 1996; Thagard, 1992; Thagard & Millgram,1995). All of these models perform parallel constraintsatisfaction using artificial neurons called units to cor-respond to mental representations such as concepts andpropositions; excitatory and inhibitory links betweenthese units correspond to constraints between repre-sentations. As described earlier, algorithms update theactivations of the units until the network of units set-tles, that is, all units have achieved stable activations.

HOTCO differs from previous cognitive models,however, in that each unit has a variable for the valenceof a unit as well as for its activation, as well as algo-rithms for updating valences as well as activations. Inaddition, a special unit called VALENCE, whose va-lence is always the maximum value of 1, provides inputto units corresponding to representations with an in-herent emotional value. In a standard connectionist al-gorithm, the activation of a unit j, aj, is updated accord-ing to the following equation:

aj(t+1) = aj(t)(1-d) + netj(max - aj(t)) if netj > 0,otherwise netj(aj(t) - min).

Here d is a decay parameter (say .05) that decrementseach unit at every cycle, min is a minimum activation(-1), max is maximum activation (1). Based on theweight wij between each unit i and j, we can calculatenetj , the net input to a unit, by:

netj = ∑iwijai(t).

HOTCO uses similar equations to update valences.The valence vj of a unit uj is the sum of the results ofmultiplying, for all units ui to which it is linked, the ac-tivation of ui times the valence of ui, times the weightof the link between ui and uj. The actual equation usedin HOTCO to update the valence vj of unit j is similar tothe equation for updating activations:

vj(t+1) = vj(t)(1-d) + netj(max- vj(t)) if netj > 0,netj(vj(t) - min) otherwise.

277

EMOTIONAL GESTALTS

Here d is a decay parameter (say .05) that decrementseach unit at every cycle, min is a minimum valence(-1), max is maximum valence (1). Based on the weightwij between each unit i and j, we can calculate netj, thenet valence input to a unit, by:

netj = ∑iwijvi(t)ai(t).

Updating valences is just like updating activations plusthe inclusion of a multiplicative factor for valences.What these equations do is ensure that the valence of aunit, corresponding to the emotional value of a repre-sentation, increases if there is positive valence flowinginto it from other active units.

In the original version of HOTCO, activations couldinfluence valences but not vice versa (Thagard, 2000).This reflects the normatively appropriate strategy thatbeliefs should affect emotions but not the reverse. Inreal life, however, people are subject to wishful think-ing and motivated inference where what they want in-fluences what they believe to be true (Kunda, 1990).Support for the influence of desires on beliefs comesfrom recent findings in behavioral decision theory con-cerning cases where the outcomes of decisions or thetargets of judgments are affectively rich (Finucane,Alhakami, Slovic, & Johnson, 2000; Rottenstreich &Hsee, 2001). These findings contradict both expectedutility theory and prospect theory, which assume thatprobability and utilities are independent. Forgas (1995)reviews the role of affective states in social judgments,explaining motivated processing effects in his AffectInfusion Model (AIM), which specifies how and whenthe valence of a mood state may infuse judgments.

Accordingly, HOTCO 2 allows units to have theiractivations influenced by both input activations and in-put valences (Thagard, forthcoming). The basic equa-tion for updating activations is the same as the standardone, but the net input to a unit’s activation is defined bya combination of activations and valences:

netj = ∑iwijai(t) + ∑iwijvi(t)ai(t).

The first summation is the standard one that makes theinput to a unit’s activation a function of the activationsof all the units linked to it, while the second is the va-lence net input defined above. Together they make ac-ceptance depend in part on desirability. These equa-tions are implemented in a LISP program available onthe Web at http://cogsci.uwaterloo.ca.

HOTCO has been used to simulate various psycho-logical phenomena, including trust and juror reason-ing. Figure 1 shows an abstract network in which acombination of emotion input and evidence input gen-erate mood by means of an overall assessment of co-herence. In the case of Gordon, the new evidence thathis paper has been rejected produces a temporary reap-

praisal of his situation and himself, dramatically alter-ing his mood.

For social psychologists, the most interestingHOTCO 2 simulation involves studies of stereotypeactivation reported by Sinclair and Kunda (1999).They found that participants who were praised by aBlack individual tended to inhibit the negative Blackstereotype, while participants who were criticized bya Black individual tended to apply the negative ste-reotype to him and rate him as incompetent. Accord-ing to Sinclair and Kunda, the participants motivationto protect their positive views of themselves causedthem either to suppress or to activate the negativeBlack stereotype.

Figure 2 shows the structure of a simplified HOTCO2 simulation of aspects of the experiment in whichpraise and criticism produced very different evaluationsof the individualwhoprovided them(Thagard, inpress).Without any evidence input that the evaluation is goodor bad, the program finds equally acceptable the claimsthat the evaluator is competent or incompetent. How-ever, a positive evaluation combines with the motivationfor self-enhancement to generate positive judgments ofthe evaluator and blacks, while a negative evaluationcombines with self-enhancement to generate negativejudgments of the evaluator and blacks. In the simulationshown in Figure 2, the positive valence of the I am goodunit supports activation of the accurate-evaluation unit,which activates the competent-manager unit and sup-presses the black stereotype. HOTCO 2 thus shows howthinking can be biased by emotional attachment to goalssuch as self-enhancement

It should be evident from this example and theprevious discussion that HOTCO models a dynamicalsystem. The variables are activations and valences ofthe various units, and the equations that define a state

278

THAGARD & NERB

Figure 1. Mood changes affected by emotional coherence. Thicklines are valence links and thin lines are cognitive links. Valencesspread from emotion input, while activations spreads from evi-dence input. Appraisals and moods are influenced by both va-lences and activations. From Coherence in thought and action byP. Thagard, 2000, p. 209. Copyright 2000 by MIT Press. Re-printed with permission.

space for the system are the equations for updatingactivations and valences stated above. The equationsare nonlinear in the mathematical sense that the vari-ables are related multiplicatively, and the system isalso nonlinear in the metaphorical sense that there aremany interactive feedback influences. The HOTCOnetwork settles into very different states dependingon changes in the input representing a positive ornegative evaluation. Similarly, the experimental par-ticipants developed a very different emotional gestaltdepending on whether the Black professional praisedor criticized them.

Although HOTCO can model some interesting psy-chological phenomena, it is far from providing a com-plete account of the range of emotions discussed by ap-praisal theory. HOTCO does not differentiate betweenspecific positive or negative emotions such as sadnessand anger directed at particular objects or situations,although it does compute a kind of global happinessand sadness based on measures of constraint satisfac-tion (Thagard, 2000). The main specific emotion thatHOTCO models is surprise, which is generated whenunits undergo a substantial change in degree of activa-tion. However, another recent model allows for greaterdifferentiation of emotions.

ITERA

Nerb and Spada (2001) present a computational ac-count of how media information about environmentalproblems influences cognition and emotion. When peo-ple hear about an environmental accident, they may re-spond with a variety of emotions such as sadness and an-ger. Following the appraisal theory of emotions, NerbandSpadahypothesized thatanegativeeventwill lead tosadness if it is caused by situational forces outside ofanyone’s control. But an environmental accident willlead to anger if someone is responsible for the negative

event. If people see themselves as responsible for thenegative event, then they feel shame; but if people seethemselves as responsible for a positive event, they feelpride.NerbandSpada(2001)showhowdeterminantsofresponsibility such as agency, controllability of thecause, motive of the agent, and knowledge about possi-ble negative consequences can be incorporated into acoherence network called ITERA (Intuitive Thinking inEnvironmental Risk Appraisal).

ITERA is an extension of the impression-forma-tion model of Kunda and Thagard (1996). The maininnovation in ITERA is the addition of units corre-sponding to emotions such as anger and sadness, asshown in Figure 3. ITERA is given input concerningwhether or not an accident was observed to involvedamage, human agency, controllability, and other fac-tors. It then predicts a reaction of sadness or angerdepending on their overall coherence with the ob-served features of the accident. This reaction can bethought of as a kind of emotional gestalt that summa-rizes all the available information.

In three studies investigating the evaluation of en-vironmental accidents, Nerb and Spada (2001) com-pared the performance of ITERA with people’s reac-tions. In their empirical experiments and theirsimulations experiments, the authors varied the de-gree of responsibility for an accident by manipulatingthe determining factors of damage, human agency,controllability, higher goal, and knowledge (cf.Weiner, 1995). In the domain of environmental prob-lems, a higher goal may be given if the accident hap-pened as a consequence of an action that was meantto achieve something that had a positive outcome,such as an increase in the overall benefit for society.Knowledge reflects whether agents knew in advancethat there is a possible contingency between their ac-tion and threats to the environment. In the ITERAmodel, information that is known about the accident

279

EMOTIONAL GESTALTS

Figure 2. Evidential and valence associations leading to the motivated inhibition of the negative black stereotype. Solid lines are excit-atory links and dashed lines are inhibitory. From “Why wasn’t O.J. convicted? Emotional coherence in legal inference, in press, Cog-nition and Emotion. Copyright 2002 by Psychology Press Ltd., Hove, UK. Reprinted with permission.

has a special status. Nodes representing such infor-mation have links to one of the two special nodes,OBSERVED+ or OBSERVED-; these special nodeshave fixed maximal or minimal activation. A link toOBSERVED+ reflects that the corresponding deter-minant is given, a link to OBSERVED- means that itis not given. No link to one of the specials nodes rep-resents that nothing is known about the determinant.By manipulating links to theses nodes, different ex-perimental settings were realized; see Figure 3 for aconcrete example of an accident scenario.

Because the determinant-emotion links arebidirectional in ITERA, there is a feedback relation-ship between determinants and emotions. These feed-back relations allow for interactions among the activa-tions of nodes for cognitive determinants. Hence thefinal activation of a node can be different across simu-lation experiments in which the inputs to other nodesare varied, so that the model predicts that people willconstrue an aspect of a situation differently dependingon other aspects of the situation. Nerb and Spada calledthese types of model predictions coherence effects. Forinstance, ITERA predicts that manipulating the con-trollability node will produce coherent patterns for hu-man agency, higher goal, and knowledge. In ITERA, acontrollable cause produces high activation values forhuman agency and knowledge but low activation forhigher goal, whereas an uncontrollable cause leads tolow activation values for human agency and knowl-edge but high activation for higher goal. Note that thereare no direct links between the cognitive determinants.The bidirectional spreading of activation within theparallel constraint satisfaction network of ITERA issufficient for producing this coherent covariation of thecognitive determinants.

Overall, ITERA produced a good fit for the modelpredictions with data for anger and the intention to

boycott the transgressor. In particular, the predictedcoherence pattern among appraisal criteria were con-firmed by empirical evidence (see Nerb, Spada, &Lay, 2001, for more empirical evidence for themodel). Such interaction effects among appraisal cri-teria are compatible with existing appraisal theoriesand are also supported by recent empirical findings(Lazarus, 1991; Lerner & Keltner, 2000, 2001). Inter-actions among appraisal criteria tend to be ignored byother types of computational models and bynon-computational appraisal models. For instance,rule-based appraisal models that realize the cogni-tion-emotion relationship as a set of if-then associa-tions do not capture the interaction effects among ap-praisal criteria (e.g., Scherer, 1993). ITERA accountsfor such affective coherence effects among appraisalcriteria by using bidirectional links for the cogni-tion-emotion relationship.

Unlike HOTCO, ITERA does not incorporate vari-ables for valence as well as activation, and it lacks al-gorithms for global calculations of coherence. But it ismore psychologically realistic with respect to the dif-ferentiation of particular emotions such as sadness andanger, and Nerb is now working on a synthesis ofHOTCO and ITERA intended to combine the best fea-tures of both. Another connectionist model of emo-tional cognition has been produced by Mischel andShoda (1995), as part of their cognitive-affective sys-tem of personality.

Therearemanyotherpossibilities for futuredevelop-mentsofcomputationalmodelsof thedynamicsofemo-tion and cognition. Wagar and Thagard (in press) de-scribeamuchmoreneurologically realisticmodelof theinteractions between emotion and cognition. It is morerealistic than HOTCO both with respect to individualneurons and with respect to the anatomical organizationof the brain. The new model uses distributed representa-

280

THAGARD & NERB

Figure 3. ITERA network for emotional reactions to environmental accidents. Solid lines are excitatory links, and dashed lines are in-hibitory links. This example represents a situation in which a media report states that there is damage caused by human agency thatcould not have been controlled. From “Evaluation of environmental problems: A coherence model of cognition and emotion,” by J. Nerband H. Spada, 2001, Cognition and Emotion, p. 528. Copyright 2001 by Psychology Press Ltd., Hove, UK. Reprinted with permission.

tions, which spread information over multiple artificialneurons, rather than the localist ones in HOTCO andITERA, which use a single neuron to stand for a conceptor proposition. In addition, the artificial neurons inWagar and Thagard’s new model behave by spiking inthemannerof realneurons, rather thansimplyspreadingactivation. Moreover, they are organized in modulescorresponding to human neuroanatomy, including thehippocampus, neocortex, amygdala, and nucleusaccumbens.Theresult is intendedtobeamodel thatcap-tures much more of the dynamic activities of the brainthan previous connectionist models of emotional cogni-tion. Wagar and Thagard (in press) simulate some of thefascinating phenomena discussed by Damasio (1994),especially the decision-making deficits found in pa-tients such as Phineas Gage who have brain damage thatdisrupts theflowof informationbetweenareas responsi-ble for reasoning (the neocortex) and emotions (theamygdala).

Much remains to be done to understand long termemotional change. Psychotherapy can take months oreven years to change the emotional tendencies of an in-dividual by helping people to revise their cognitionsand emotions. A marriage that begins with mutual lovecan dissolve into anger and even hatred. The terroristleader Osama bin Laden began with a dislike for Amer-icans that eventually turned into an intense hatred. Adynamical theory of emotion should deal with longterm emotional changes as well as the sudden transi-tions discussed in this article.

Another direction for future research on the dynam-ics of emotion is to expand models to capture group in-teractions. Thagard (2000, ch. 7) developed a theory ofconsensus resulting from communication betweenmembers of a scientific community who disagree aboutwhat theory provides the best explanation of the avail-able evidence. He has run computer simulations inwhich up to 60 scientists, each of whom makes coher-ence-based decisions, reach agreement on disputed sci-entific issues. Future work will develop simulations inwhich group decision making involves emotions as wellasevidence.Thiswill requireextensionofexistingmod-els of emotional coherence and consensus to apply togroup decisions in which individuals communicate notonly theirbeliefs andgoalsbut also their emotional eval-uations of competing options.

Conclusion

Although the theoretical vocabulary and computa-tional modeling described in this article may seem unfa-miliar to social psychologists, the basic ideas are similarto those of some of the classic theories in the field.Festinger’s (1957) notion of cognitive dissonance canbe accounted for in terms of parallel constraint satisfac-tion (Shultz & Lepper, 1996); and it is possible with

HOTCO to incorporate the affective dimension thatlater theorists found to be crucial to dissonance (Cooper& Fazio, 1984). Kunda and Thagard (1996) describehow parallel constraint satisfaction can be used to un-derstand ideas about impression formation that Solo-mon Asch developed from Gestalt psychology.

Emotional coherence and the kind of computationalmodel discussed in this article can also be applied toMcGuire’s (1999) dynamic model of thought systems.McGuire’s theory describes thought systems as consist-ing of propositions with two attributes: desirability (anevaluation dimension) and likelihood of occurrence (anexpectancy dimension). These dimensions express howmuch the content of a proposition is liked and how muchit is believed. The thought system is seen as dynamic sothat a change that is directly induced in one part of thesystem results in compensatory adjustments in remoteparts of the system. Together those assumptions implythat a person’s evaluative and expectancy judgments ona topic will tend to be assimilated toward one another.McGuire (1999) postulated “that causality flows in bothdirections, reflecting both a ‘wishful thinking’tendencysuch that people bring their expectations in line withtheir desires, and a ‘rationalization’ tendency such thatthey bring desires in line with expectations” (p. 190).Emotional coherence models such as HOTCO providemechanisms for both these tendencies, as activationscorresponding to likelihood interact with valences cor-responding to desirability.

Some proponents of a dynamical systems approachto understanding the mind have seen it as a radical al-ternative to symbolic and connectionist models thathave been predominant in cognitive science (vanGelder, 1995). Our discussion has shown that this is amistake: connectionist systems are one important kindof dynamical system, and they are arguably the onesbest suited for explaining psychological phenomena,including ones involving emotion.

In sum, computational models based on parallelconstraint satisfaction, with constraints and represen-tations that are affective as well as cognitive, havemuch to contribute to a dynamical perspective on so-cial psychology. We have shown how connectionistsystems such as HOTCO and ITERA can be used tostudy the dynamics of emotional cognition at a con-crete and not just a metaphorical level. In particular,they can model the generation and shifting of emo-tional gestalts.

References

Abelson, R. (1963). Computer simulation of “hot” cognition. In S.Tomkins & S. Messick (Eds.), Computer simulation of person-ality (pp. 277–298). New York: Wiley and Sons.

Bower, G. H. (1981). Mood and memory. American Psychologist,36, 129–148.

281

EMOTIONAL GESTALTS

Cooper, J., & Fazio, R. H. (1984). A new look at dissonance theory.In L. Berkowitz (Ed.), Advances in experimental social psy-chology (Vol. 17). New York: Academic.

Damasio, A. R. (1994). Descartes’ error. New York: G. P. Putnam’sSons.

Festinger, L. (1957). A theory of cognitive dissonance. Stanford, CA:Stanford University Press.

Finucane, M. L., Alhakami, A. S., Slovic, P., & Johnson, S. M.(2000). The affect heuristic in judgements of risks and benefit.Behavioral Decision Making, 13, 1–17.

Forgas, J. P. (1995). Mood and judgment: The Affect Infusion Model(AIM). Psychological Bulletin, 117, 39–66.

Holyoak, K. J., & Spellman, B. A. (1993). Thinking. Annual Reviewof Psychology, 44, 265–315.

Holyoak, K. J., & Thagard, P. (1989). Analogical mapping by con-straint satisfaction. Cognitive Science, 13, 295–355.

Kunda, Z. (1990). The case for motivated inference. PsychologicalBulletin, 108, 480–498.

Kunda, Z., & Thagard, P. (1996). Forming impressions from stereo-types, traits, and behaviors: A parallel-constraint-satisfactiontheory. Psychological Review, 103, 284–308.

Lazarus, R. S. (1991). Cognition and motivation in emotion. Ameri-can Psychologist, 46, 352–267.

Lerner, J. S., & Keltner, D. (2000). Beyond valence: Toward a modelof emotion-specific influences on judgement and choice. Cog-nition and Emotion, 14, 473–493.

Lerner, J. S., & Keltner, D. (2001). Fear, anger, and risk. Journal ofPersonality and Social Psychology, 81, 146–159.

Lewin, K. (1951). Field theory in social science. New York: Harper& Row.

McGuire, W. J. (1999). Constructing social psychology: Creativeand critical processes. New York: Cambridge UniversityPress.

Mischel, W., & Shoda, Y. (1995). A cognitive-affective system the-ory of personality: Reconceptualizing situations, dispositions,dynamics, and the invariance in personality structure. Psycho-logical Review, 102, 246–268.

Nerb, J., & Spada, H. (2001). Evaluation of environmental problems:A coherence model of cognition and emotion. Cognition andEmotion, 15, 521–551.

Nerb, J., Spada, H., & Lay, K. (2001). Environmental risk in the me-dia: Modeling the reactions of the audience. Research in SocialProblems and Public Policy, 9, 57–85.

Port, R., & van Gelder, T. (Eds.). (1995). Mind as motion: Explora-tions in the dynamics of cognition. Cambridge, MA: MIT Press.

Read, S. J., Vanman, E. J., & Miller, L. C. (1997). Connectionist, par-allel constraint satisfaction, and Gestalt principles: (Re)Intro-ducing cognitive dynamics to social psychology. Personalityand Social Psychology Review, 1, 26–53.

Rottenstreich, Y., & Hsee, C. K. (2001). Money, kisses, and electricshocks: On the affective psychology of risk. Psychological Sci-ence, 12, 185–190.

Rumelhart, D. E., & McClelland, J. L. (Eds.). (1986). Parallel dis-tributed processing: Explorations in the microstructure of cog-nition. Cambridge MA: MIT Press/Bradford Books.

Scherer, K. R. (1993). Studying the emotion-antecedent appraisalprocess: An expert system approach. Cognition and Emotion, 7,325-355.

Scherer, K. R., Schorr, A., & Johnstone, T. (2001). Appraisal pro-cesses in emotion. New York: Oxford University Press.

Shultz, T. R., & Lepper, M. R. (1996). Cognitive dissonance reductionasconstraint satisfaction.PsychologicalReview,103,219-240.

Sinclair, L., & Kunda, Z. (1999). Reactions to a Black professional:Motivated inhibition and activation of conflicting stereotypes.Journal of Personality and Social Psychology, 77, 885-904.

Thagard, P. (1992). Conceptual revolutions. Princeton: PrincetonUniversity Press.

Thagard, P. (1996). Mind: Introduction to cognitive science. Cam-bridge, MA: MIT Press.

Thagard, P. (2000). Coherence in thought and action. Cambridge,MA: MIT Press.

Thagard, P. (in press). Why wasn’t O. J. convicted? Emotional coher-ence in legal inference. Cognition and Emotion.

Thagard, P., & Millgram, E. (1995). Inference to the best plan: A co-herence theory of decision. In A. Ram & D. B. Leake (Eds.),Goal-driven learning (pp. 439-454). Cambridge, MA: MITPress.

Thelen, E., & Smith, L. B. (1994). A dynamic systems approach to thedevelopment of cognition and action. Cambridge, MA: MITPress.

Vallacher, R. R., Read, S. J., & Nowak, A. (2002). The dynamicalperspective in personality and social psychology. Personalityand Social Psychology Review, 6, 264–273.

van Gelder, T. (1995). What might cognition be, if not computation?Journal of Philosophy, 92, 345-381.

Wagar, B., & Thagard, P. (in press). Using computational neurosci-ence to investigate the neural correlates of cognitive-affective in-tegration during covert decision making. Brain and Cognition.

Weiner, B. (1995). Judgements of responsibility: A foundation for atheory of social conduct. New York, NY: Guilford Press.

282

THAGARD & NERB