Designing Personality: Cognitive Architectures and Beyond

Modeling personality is important both for understanding human traits, and for designing artificial systems. Both aims are challenging. Studies of human personality suggest that personality traits such as extraversion and neuroticism have multiple correlates relating to many qualitatively different aspects of neural and cognitive functioning. The tri-level explanatory framework of cognitive science provides a basis for personality description. Traits are expressed as biases in neural functioning, in the functional architecture supporting symbolic processing, and in high-level selfknowledge and motivation. Furthermore, traits relate to multiple biases of each type, so that traits are distributed across and within levels. Such a description supports finegrained modeling of personality. A similar approach may be applied to modeling transient states such as emotions. However, the descriptive picture is incomplete. Personality also has a dynamic, adaptive aspect, as the system learns to fulfill its goals by acquiring contextualized skills appropriate to the external environment. Thus, in humans, observed differences in architectural parameters must be interpreted adaptively. In designing artifacts, there may be advantages to allowing personality to emerge in part from interactions between the artifact and the environments within which it is intended to function. Personality and Cognitive Architecture Personality theory needs cognitive architecture – and studies of cognitive architecture need personality. Traditionally, the theory of stable personality traits has been derived from biological psychology. Theorists such as Eysenck (1967) and Gray (1991) have proposed that individual differences in brain function generate expressions of personality ranging from psychophysiological response to social learning. For several decades, the search has been on for key brain systems, controlling arousal or motivation, that are the locus for the major personality traits such as extraversion and neuroticism. It is now beyond dispute that such traits do indeed have biological bases, as evidenced by their partial heritability, their sensitivity to brain lesions, and Copyright © 2004, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. psychophysiological correlates of traits (Zuckerman, 1991). However, cognitive psychological studies show the weaknesses of the traditional biological models in explaining the behavioral correlates of traits. Experimental tests of predictions derived from the Eysenck and Gray models have met with limited success. Even when predictions are confirmed, effect sizes are normally modest, and indices of psychophysiological change fail to mediate effects of personality on performance (Matthews & Amelang, 1993). By contrast, effects of personality are often predictably moderated by the information processing demands of the task. Task factors well-known to cognitive psychologists such as working memory load, stimulus onset asynchronies in priming, and demands on attention influence the direction and magnitude and of trait effects (Matthews, 1997a). The behavioral consequences of personality cannot be explained adequately in terms of gross qualities such as arousability and conditionability. Instead, a more fine-grained analysis of the effects of personality on information-processing is required. In consequence, linking personality to individual differences in parameters of the cognitive architecture is critical. The argument that the study of cognitive architectures requires attention to personality factors is less well-known. Recent interest in personality in artificial intelligence derives in part from issues raised by the design of emotionally intelligent artifacts. The display of emotionally appropriate responses to user input may be an important design feature of such systems, facilitating naturalistic communication. For example, Ball (2003) describes an outline architecture for a conversational interface that might be applied, initially, to domain limited applications such as buying a travel ticket from an artificial travel agent. The architecture includes modules that assess emotion and personality from the sensory cues provided by the user. A ‘policy module’ uses this information to judge an appropriate emotional response, which influences the behavior of the agent via a Bayesian network. Ortony (2003) views personality as a potential driver of behavior in affective artifacts. The design of believable artifacts requires stability and coherence in behavior, which requires that personality traits should be built into the architecture. Ortony suggests that a principled approach to the design issues may be derived from standard trait models, such as the Five Factor Model (FFM: Costa & McCrae, 1992). Each trait may be linked to a set of key parameters able to generate (in interaction with environmental input) a range of trait-characteristic states and behaviors. Thus, to build an artifact that is friendly, those behaviors it produces that convey friendliness should be generated by biases in some rational system for handling emotion. For example, the system could be biased to appraise the human user as well-intentioned, to follow goals of increasing user comfort, and to prioritize inputs indicating discomfort. Such ‘motivated’ friendliness is likely to be more believable, and hence more effective, than simply having the system emit more frequent ‘friendly’ responses irrespective of context. In this article, I will summarize what can be learnt from human experimental studies of personality, affect and performance that may facilitate designing personality into cognitive architectures. I will focus primarily on the use of architectures to simulate human behavior, but I will also refer to implications for design of artifacts. Next, I will outline the key trait and state constructs that are open to modeling. I will go on to argue that modeling individual differences in parameters of the architecture is necessary but not sufficient. Traits and states have important effects on neural functioning and intention, as well as on computation, that call for a multi-level approach. Pylyshyn’s (1984, 1999) tri-level cognitive science framework provides a systematic approach to mapping the various attributes of traits and states. I will argue that these attributes are distributed both within and across levels of explanation. Traits are supported by a multitude of typically small biases at various levels of abstraction from the neural substrate. Finally, I will argue that fine-grained descriptive accounts should be integrated with a dynamic perspective that relates traits to individual differences in adaptation. Traits may correspond to environmental affordances, that, over time, build congruence between basic parameters of the architecture, acquired skills and ongoing interaction with environmental demands and opportunities. Key Constructs: Traits and States There is a plethora of dispositional traits and transient states that may be related to properties of the cognitive architecture. Although traits are defined in terms of temporal stability, personality changes over time, and the roles of maturation and learning in personality change are important topics of study. Here, I will focus primarily on short-term behavioral change, so that traits may be treated as ‘fixed’. One of the most useful trait frameworks has proved to be the FFM. Each of the five traits represents a continuum contrasting qualities that are polar opposites, as shown simplistically in Table 1 (see Matthews, Deary & Whiteman, 2003, for a review). Also important are more narrowly-defined traits, that sit beneath the Big Five in hierarchical models of personality, such as impulsivity, sensation-seeking, optimism-pessimism and many others. Typically, researchers are concerned with modeling a single trait only, but, as many behaviors are influenced by multiple traits, multiple-trait modeling may become increasingly important. Table 1. Examples of qualities defining the high and low poles of the ‘Big Five’ traits. Trait High Pole Low Pole Extraversion Sociable Withdrawn Neuroticism Vulnerable to Stress Emotionally stable Conscientiousness Hardworking Careless Agreeableness Sympathetic Uncaring Openness Imaginative Practical In each case, we may seek to relate the trait to fundamental parameters of the cognitive architecture, such as memory spaces, connection strengths, and speed of processes, representing the trait as a collection of structurally independent biases. However, traits also bias the acquisition of skills and knowledge, so that, for example, the content of memory as well as its formal operating parameters may vary according to personality. By contrast with traits, research on transient states offers a hazier picture of dimensionality of constructs, but more sophisticated cognitive models that draw on emotion theory. I will confine discussion here to subjective states, including not just affective states such as moods and emotions, but also cognitive states, such as worry, and motivational states such as apathy. Most of the psychometric work has been conducted on mood states or basic affects, with reasonable agreement in favor of either two dimensions of positive and negative affect, or three dimensions of energy, tension and pleasure (Schimmack & Grob, 2000). In recent work, my colleagues and I (Matthews, Campbell et al., 2002) have discriminated three state dimensions that integrate affect with cognition and emotion: task engagement (energy, motivation and concentration), distress (tension, displeasure, low confidence) and worry (self-focus, intrusive thoughts, low self-esteem). The main areas of research that relate states to performance and information-processing center on state anxiety, overall mood (often related to cognitive bias) and fatigue. As with traits, states have complex effects dependent on a host of moderating factors. Several rather different types of explanation for associations between states and performance have been advanced (Matthews, Zeidner & Roberts, 2002): • States may be associated with temporary changes in parameters of the architecture. Such changes may be described as specific local effects, such as prioritization of threat processing in anxiety, or as a system-wide reconfiguration geared to the motivational context (Oatley & Johnson-Laird, 1996), so that changes in operational parameters are functionally linked. • States may relate to changes in strategy, for example, in voluntary allocation of effort. In this case, the parameters of the architecture may remain the same, but the usage of the functionality the architecture provides is different. • States may operate through representation rather than computation (cf., the idea of ‘emotion-asinformation’). In Bower’s (1981) classic semantic network model, states are represented as network nodes whose activation follows the same rules as the nodes for other types of concept. States may influence the contents of memory; for example, effects of worry on performance appears to reflect the intrusion into working memory of selfreferent cognitions. • States may represent targets for self-regulation; i.e., as drivers of voluntary attempts at mood-regulation (cf. the idea of ‘emotion-as-motivation’) . People differ in their metacognitions of the importance of changing moods and cognitive states, and in their strategy preferences. Some trait and state effects may be inter-related, whereas others are separate. Often, effects of traits are said to be mediated by states. Typically, extraversion is linked to positive affect, whereas neuroticism is related to negative affect (Watson, 2000). Thus, extraverts and introverts may differ in how they process information because the trait predisposes differing affective states, which are the more proximal influence on processing. However, trait effects can also be demonstrated with state factors statistically controlled, indicating that some trait effects are not mediated by individual differences in states (Matthews & Harley, 1993). Such ‘direct’ effects may indicate that traits influence parameters of the architecture that are insensitive to state effects. These effects may be a consequence of learning processes operating over many years. Architecture and Beyond: A Multi-Level Explanatory Framework Cognitive Architectures and Their Limitations Extensive empirical research indicates that, typically, any given trait or state (or trait × state interaction) correlates with multiple attributes of performance (Matthews et al., 2003). By analogy with cognitive stress research (Hockey, 1984), effects of trait and state factors on performance may be described by a cognitive patterning, which describes how the factor influences a set of performance indicators chosen to reflect key processing functions, such as attentional selectivity and working memory capacity. Such a description leaves open the issue of how each empirical effect of the factor may explained. Certainly, it can be shown that some effects can be modeled computationally, thereby relating the trait to individual differences in parameters of the architecture (e.g., Matthews & Harley, 1993). However, a purely architecture-based approach also has limitations of its own. First, effects on performance may be either ‘structural’, changing the operation of some basic processing function, or ‘strategic’, changing the person’s voluntary choice of task goals and subgoals (Hockey, 1984). For example, anxiety is associated with increased attentional selectivity, but it is unclear whether the effect relates to some ‘automatic’ narrowing of an attentional spotlight, or to a voluntary decision to concentrate attention on salient, potentially threatening stimuli (see Eysenck, 1992). Second, while effects of personality and emotion on strategy may be mapped descriptively, it remains unclear why there are individual differences in taskand self-referent goals. Third, the cognitive patterning approach risks losing sight of the core attributes of personality traits. For example, it is unclear how biases in information-processing generate the attributes such as sociability, impulsivity and assertiveness that are central to extraversion. Fourth – in line with both interactionist theories of personality and situated approaches to emotion – a description of architecture that neglects its functionality within some external environment cannot capture individual differences in dynamic person-situation interaction. The Tri-level Explanatory Framework One solution to these difficulties is to develop multileveled conceptions of traits and emotions. According to the ‘classical theory’ of cognitive science (Newell, 1982; Pylyshyn, 1984, 1999), cognitive phenomena are open to three complementary, types of explanation. The first is the biological level, which refers to the neural ‘hardware’ supporting processing. Individual differences in performance might reflect variation in brain functioning, as proposed by biological personality theories. The second level of explanation is described by Pylyshyn (1984) as the symbolic level, referring to the formalized computational operations which constitute the ‘software’ of the mind, and the software facilities such as memory space and communication channels which support processing. I have suggested elsewhere (Matthews, 1997b, 2000) that in personality research this level of explanation might be reconceptualized as a cognitive-architectural level, in order to accommodate sub-symbolic, connectionist models. As in the typical cognitive-psychological model, the aim is to identify the specific processing components which mediate personality effects. The third level of explanation is labeled the semantic level by Pylyshyn (1984), in that it refers to the personal meaning of the otherwise arbitrary processing codes, and the knowledge level by Newell (1982), because it refers to the person’s knowledge of how to obtain personal goals. More generally, it explains behavior on the basis of 1 I will leave aside the issue of whether higher levels are ultimately reducible to the biological level (see Matthews, 2000) intentions, motivations and personal meaning. It is compatible with cognitive stress models, which attribute stress symptoms to the status of self-regulative plans for goal-satisfaction (Wells & Matthews, 1994). Thus, accounts of personality and emotion expressed in terms of the cognitive architecture are necessary but not sufficient. We need to look downwards from the cognitive architecture to its neural foundations, the appropriate level for explaining responses closely coupled with the neural substrate. We also need to look upwards to the knowledge level to understand why personality and affect may be correlated with personal goals and strategy choice. Personality Traits as Distributed Constructs Extraversion and Neuroticism How can the multi-level approach be realized in practice? Table 2 illustrates some of the correlates of extraversionintroversion, allocating each of several of the more robust empirical findings to one of the three levels (see Matthews, 1997a, for a more detailed review). These allocations are tentative because researchers often neglect to probe mechanisms in detail, but some modeling studies are at least suggestive of where specific effects should be placed (e.g., Matthews & Harley, 1993). The key feature of the table is that personality effects are distributed at every level. At the biological level, Zuckerman (1991) emphasizes that there is no isomorphism between traits and specific brain systems; instead, multiple systems may underpin each trait. Consistent with this principle, a review of psychophysiological studies (Matthews & Gilliland, 1999) concluded that there are at least two reliable sets of correlations of extraversion: one set relating to corticoreticular arousability, and a second set associated with dopaminergic reward systems and motor responsivity. Effect sizes for these correlations are typically modest. Table 2. Differing explanations for empirical correlates of extraversion: some examples. Level of explanation Empirical correlates Biological Conditioning to reward Low cortical arousability Motoneuronal excitability (low) CognitiveVerbal divided attention architectural Sustained attention (poor) Fluency of speech production Knowledge Social motivations and interests Appraisal of events as challenging Coping using task-focused strategies At the level of cognitive architecture (and admitting that some effects may in fact be strategic), extraverts process information more efficiently when the task involves multiple channels of information (especially verbal information), such as dual-task performance and resisting distraction. Extraverts also exhibit superior speech production abilities and faster memory retrieval. Introverts, however, are superior at sustaining attention over time in simple vigilance paradigms. At the knowledge level are performance effects associated with strategy change such as setting a response criterion and maintaining focus during reflective problem-solving: extraverts tend to be more impulsive. More generally, extraverts and introverts differ in the meanings they assign to events, and the goals and coping strategies they adopt in challenging situations. Table 3 shows a comparable analysis of correlates of neuroticism and trait anxiety (Matthews, Derryberry & Siegle, 2000). Like extraversion, neuroticism is distributed across neural processes (e.g. sensitivity to punishment signals), basic information-processing operations (e.g., disengagement of attention from sources of threat) and high-level self-regulation and motives (e.g., biases towards threat appraisal and emotion-focus coping). In principle, multi-level analyses could be applied to other Big Five Traits. Future research might relate conscientiousness to processes governing effort and task commitment, agreeableness to social cognitions, and openness to the intellect (Matthews et al., 2003). Table 3. Differing explanations for empirical correlates of neuroticism: some examples. Level of explanation Empirical correlates Biological Conditioning to punishment Immune response (weak) Startle response to fear stimuli CognitiveAttentional bias to threat architectural Selective memory for negative events Negative bias in judgement Knowledge Self-protective motives Appraisal of events as threatening Coping using emotion-focused strategies Levels of Affective Functioning How does affect fit into the multi-leveled picture? There are two issues here: why emotions (and other states) influence cognition and performance, and why personality relates to individual differences in emotions. As with personality, it is likely that emotion operates at all three levels of the cognitive science framework (Matthews et al., 2000). The neurological foundations of emotion are well-known, although it is regrettable that neuroscientists have tended to neglect relationships between individual differences in brain systems and in performance. Fear-conditioning might be a good example of a process best understood neurologically, although it is likely that expectancies and higher-level cognitive processing also play a role. As previously noted, emotions may also be associated with specific changes in the architecture, and with information content. The knowledge-based perspective on emotion is also familiar from studies of strategy change in emotional states, and from cognitive models of stress and self-regulation. I will set aside the biological level here, so as to provide two brief examples of how research may draw on explanations based on both cognitive architecture and on self-knowledge. The first example concerns the robust association found between energetic arousal and efficiency of performance of attentionally demanding tasks, such as signal detection and controlled visual search tasks (reviewed by Matthews, 1997a). Effects of energetic arousal are reliably moderated by a variety of task demand factors, such that energy is related only to attentionallydemanding tasks. Hence, energetic mood may index attentional resource availability. Although energy primarily appears as an index of the cognitive architecture, it also correlates with task motivation, effort, challenge appraisal and task-focused coping (Matthews, Campbell et al., 2002). This broader syndrome of task engagement may be understood at the knowledge level as relating to the person’s understanding of the environment as calling for the application of systematic, goal-directed effort. Another example concerns anxiety and negative affect. Wells and Matthews (1994) proposed that such states are often generated by dysfunctional self-referent executive processing, associated with exaggerated attention to selfdiscrepancies and intrusive thoughts. This account of anxiety and depression proposes an architecture, but explains their behavioral consequences primarily in terms of how the anxious person deploys the architecture, rather than in terms of changes in parameters of the architecture (although these may occur simultaneously). For example, although threat-sensitivity in anxiety is often considered to operate ‘automatically’, data suggest that attentional bias is sensitive to expectancies and strategy. Such bias may in fact be a consequence of a strategy for handling threat by ‘hypervigilance’ for potential dangers. Matthews and Harley (1996) modeled attentional bias by relating anxiety to activation of ‘task-demand’ units that modulate the operation of a parallel distributed processing (PDP) network. The model provides an architecture-level description of the effect, but explaining why anxiety relates to variation in task-demand activation requires the knowledge level. Anxious individuals interpret the world and their own place in it differently to those low in anxiety, emphasizing the likelihood of personal danger (social or physical), and the utility of hypervigilance as a means of coping. The second issue is the relationship between personality and affect. Extraverts’ tendencies to positive affect are often attributed to a brain reward system, whereas the association between neuroticism and negative affect is said to be mediated by a punishment system (e.g., Watson, 2000). However, there is rather little evidence linking personality-mood covariation directly to neural bases (Matthews & Gilliland, 1999). My suggestion is that multiple mechanisms may support the associations. For example, in neuroticism and anxiety, informationprocessing biases that increase the salience of threat may contribute to higher levels of negative affect. So too may biases in self-knowledge, appraisal and coping, as shown directly in studies of task-induced stress (Matthews, Campbell et al., 2002). These constellations of change in neural, computational and self-regulative processes associated with transient states may partly mediate the effects of personality traits on cognition and performance. For example, the role of state anxiety in mediating effects of trait anxiety is wellknown. However, some effects appear to be unique to traits. Because of the stability of traits they become associated over time with individual differences in acquired skills, such as social skills in the case of extraverts, and skills for early recognition of danger in the case of neurotics (Matthews et al., 2003). Skill differences may influence behavior irrespective of temporary state. Conversely, if we look at true state measures, rather than aggregated ‘typical mood’ measures, much of the variance in states is not predictable from traits (instead reflecting situational factors). Thus, we cannot use traits as proxies for states in research. The Role of the Environment: Adaptation and