Analogical Mapping by Constraint Satisfaction

Keith J. Holyoak, Paul Thagard
1989 Cognitive Science  
This paper interprets emotional change as a transition in a complex dynamical system. We argue that the appropriate kind of dynamical system is one that extends recent work on how neural networks can perform parallel constraint satisfaction. Parallel processes that integrate both cognitive and affective constraints can give rise to states that we call emotional gestalts, and transitions can be understood as emotional gestalt shifts. We describe computational models that simulate such phenomena
more » ... n ways that show how dynamical and gestalt metaphors can be given a concrete realization Introduction One Tuesday morning, Professor Gordon Atwood strode energetically into his department office at Yukon University. It was a great day -the sun was shining, his family had been good natured at breakfast, and he was finished teaching for the semester. He cheerily greeted the department secretary and went to his mailbox, eagerly noticing the long-awaited letter from the Journal of Cognitive Chemistry. After opening the letter, 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." As he scanned the reviews, sadness turned to anger when Gordon realized that one of the reviewers had totally failed to understand his paper, while the other had rejected it because it did not sufficiently cite the reviewer's own work. Stomping out of the department office, Gordon chided the secretary for running out of printer paper yet again. After an unproductive morning largely spent surfing the Web, Gordon met his wife for lunch. She reminded him that he had recently had two articles accepted for publication, and the rejected article could always be published elsewhere. They chatted about their daughter's recent triumph in a ballet school performance, and made plans to see a musical that weekend. The Thai curry was delicious, and the coffee was strong. Gordon returned to work and happily updated his article for submission to a different journal. Everyone has had days like this, with transitions between different moods and emotions. A psychological theory of affect must explain both how different emotional states arise and how one state can be replaced by another one that is qualitatively very different. Affect is a natural subject for a dynamical theory that emphasizes the flow of thought and the complex interactions of emotion and cognition. Our aim in this paper is to develop such a theory of the emergence and alteration of emotional states. We proceed first by interpreting emotional change as a transition in a complex dynamical system. This metaphorical interpretation, however, is limited in its explanatory power without concrete specification of the structures and mechanisms that can give rise to emotions and emotional change. We argue that the appropriate kind of dynamical system is one that extends recent work on how neural networks can perform parallel constraint satisfaction. Parallel processes that integrate both cognitive and affective constraints can give rise to states that we call emotional gestalts, and transitions such as those experienced by Gordon can be understood as emotional gestalt shifts. Finally, we describe computational models that simulate such phenomena in ways that show how dynamical and gestalt metaphors can be given a concrete realization. Emotion as a Dynamical System At its simplest, psychological theory postulates causal relations between mental properties and behavior, for example that there is a personality trait of extraversion that leads people who have it to talk frequently with other people. Since the rise of cognitive science in the 1950s, psychological theories have increasingly postulated representational structures and computational processes that operate on the structures to produce behavior. More recently, some psychologists and philosophers have proposed that psychological theories should be analogous to theories of complex dynamical systems that have been increasingly popular in physics, biology, and other sciences (see, for example, Port and van Gelder, 1995; Thelen and Smith, 1994). Thagard (1996, ch. 11) described how dynamical systems theory can be applied to psychological phenomena by means of the following explanation schema: Explanation Target: Why do people have stable but unpredictable patterns of behavior? Explanatory Pattern Human thought is describable by a set of variables. These variables are governed by a set of nonlinear equations. These equations establish a state space that has attractors. The system described by the equations is chaotic. The existence of the attractors explains stable patterns of behavior. Multiple attractors explain abrupt phase transitions. The chaotic nature of the system explains why behavior is unpredictable. 5 It is easy to apply this explanation schema to emotions. We want to be able to explain both why people have ongoing emotions and moods, and also how they can sometimes make dramatic transitions to different emotional states. Hypothetically, we might identify a set of variables that describe environmental, bodily, and mental factors. Equations that describe the causal relations among those variables would clearly be nonlinear, in that they would require specification of complex feedback relations between the different factors. The system described by the equations would undoubtedly be chaotic, in the sense that small changes to the value of some variables could lead to very large changes in the overall system: it only took one event to dramatically change Gordon's mood. On the other hand, the emotional dynamic system does have some stability, as people maintain a cheerful or terrible mood over long periods. This stability exists because the system has a tendency to evolve into a small number of general states called attractors, and the shift from one mood to another can be described as the shift from one attractor to another. Compare the kinds of perceptual transitions that were identified by the gestalt psychologists. When you see a Necker cube or a duck-rabbit, you see more than just the lines that make up the figure. The cube flips back and forth as you see it in different gestalts, and the duck-rabbit appears to you as a duck or as a rabbit but not both. Attending to different aspects of the drawing produces a gestalt shift in which you move from one configuration to the other. In the language of dynamical systems theory, the perceptual system has two attractor states, and the gestalt shift involves a phase transition from one attractor to the other. Analogously, we might think of an emotional state as a human agency, higher goal, and knowledge. In ITERA, a controllable cause produces high activation values for human agency and knowledge but low activation for higher goal, whereas an uncontrollable cause leads to low activation values for human agency and knowledge but high activation for higher goal. Note that there are no direct links between the cognitive determinants. The bidirectional spreading of activation within the parallel constraint satisfaction network of ITERA is sufficient for producing this coherent covariation of the cognitive determinants. Overall, ITERA produced a good fit for the model predictions with data for anger and the intention to boycott the transgressor. In particular, the predicted coherence pattern among appraisal criteria were confirmed by empirical evidence (see Nerb, Spada, & Lay, 2001 , for more empirical evidence for the model). Such interaction effects among appraisal criteria are compatible with existing appraisal theories and are also supported by recent empirical findings (Lazarus, 1991; Keltner, 2000, 2001). Interactions among appraisal criteria tend to be ignored by other types of computational models and by non-computational appraisal models. For instance, rule-based appraisal models that realize the cognition-emotion relationship as a set of if-then associations do not capture the interaction effects among appraisal criteria (e.g., Scherer, 1993). ITERA accounts for such affective coherence effects among appraisal criteria by using bidirectional links for the cognition-emotion relationship. Unlike HOTCO, ITERA does not incorporate variables for valence as well as activation, and it lacks algorithms for global calculations of coherence. But it is more psychologically realistic with respect to the differentiation of particular emotions such as sadness and anger, and Nerb is now working on a synthesis of HOTCO and ITERA 20 intended to combine the best features of both. Another connectionist model of emotional cognition has been produced by Mischel and Shoda (1995) , as part of their cognitiveaffective system of personality. There are many other possibilities for future developments of computational models of the dynamics of emotion and cognition. Wagar and Thagard (forthcoming) describe a much more neurologically realistic model of the interactions between emotion and cognition. It is more realistic than HOTCO both with respect to individual neurons and with respect to the anatomical organization of the brain. The new model uses distributed representations, which spread information over multiple artificial neurons, rather than the localist ones in HOTCO and ITERA, which use a single neuron to stand for a concept or proposition. In addition, the artificial neurons in Wagar and Thagard's new model behave by spiking in the manner of real neurons, rather than simply spreading activation. Moreover, they are organized in modules corresponding to human neuroanatomy, including the hippocampus, neocortex, amygdala, and nucleus accumbens. The result is intended to be a model that captures much more of the dynamic activities of the brain than previous connectionist models of emotional cognition. Wagar and Thagard (forthcoming) simulate some of the fascinating phenomena discussed by Damasio (1994) , especially the decision-making deficits found in patients such as Phineas Gage who have brain damage that disrupts the flow of information between areas responsible for reasoning (the neocortex) and emotions (the amygdala). Much remains to be done to understand long term emotional change. Psychotherapy can take months or even years to change the emotional tendencies of an
doi:10.1207/s15516709cog1303_1 fatcat:jl4mcamgtfatngxqimhviu6gla