A model of personality should be a cognitive architecture itself

Available online at www.sciencedirect.com

www.elsevier.com/locate/cogsys

ScienceDirect

Cognitive Systems Research 29–30 (2014) 1–30

Action Editor: Vasant Honavar

Ron Sun ⇑, Nicholas Wilson

Department of Cognitive Sciences, Rensselaer Polytechnic Institute, Troy, NY 12180, USA

Received 26 September 2013; received in revised form 16 December 2013; accepted 29 January 2014Available online 19 February 2014

Abstract

This paper describes how personality may be explained by a generic, comprehensive computational cognitive architecture. We showthat a cognitive architecture by itself can serve as a generic model of personality, without any significant addition or modification. Acognitive architecture can capture the fundamental invariance within an individual in terms of behavioral inclinations as well as the inev-itable variability of behaviors. Various tests and simulations have been conducted within the cognitive architecture that show that such amodel is reasonably stable, is relatively flexible (in terms of person–situation interactions), captures some major personality traits (e.g.,the Five-Factor Model), and accounts for a variety of empirical data. The work shows the feasibility and usefulness of integratingpersonality modeling with generic computational cognitive modeling (i.e., cognitive architectures).� 2014 Elsevier B.V. All rights reserved.

Keywords: Personality; Cognitive architecture; Cognitive modeling

1. Introduction

Personality, as extensively studied in personalitypsychology, should be the result of various existing psycho-logical mechanisms and processes that have beencommonly identified, and nothing else. That is, thereshould not be any special entities, mechanisms, orprocesses of personality (as has been argued before).Cervone (2004), for example, has argued that personalityis a complex system with dynamic interactions amongmultiple processes; thus, personality should be understoodby reference to basic cognitive processes that give rise toovert behavior. Shoda and Mischel (1998), Mayer (2005),and Sun and Wilson (2011) made similar points.

A computational “cognitive architecture” should, ide-ally, include all essential psychological entities, mecha-nisms, and processes of the human mind (Sun, 2004). The

⇑ Corresponding author.E-mail address: dr.ron.sun@gmail.com (R. Sun).

notion of a cognitive architecture should ideally be closeto that of “personhood” (Pollock, 2008; Taylor, 1985).Within the cognitive architecture, the interaction amongdifferent subsystems (components and their mechanismsand processes) should be able to generate psychologicalphenomena of all kinds, which of course include personal-ity-related phenomena (Sun & Wilson, 2011).

Thus, personality, if it is a valid psychological construct,should be accounted for by a cognitive architecture, with-out any significant additions or modifications of mecha-nisms or processes within the cognitive architecture. Acognitive architecture should have essential components,mechanisms, and processes of the mind (such as variousmemory modules, inference mechanisms, learning mecha-nisms, and so on) built-in. Otherwise, it would amount toa software tool, which allows one to build whatever modelsthat one wants to build but does not sufficiently specify thearchitecture of the mind. In addition, it is desirable that amodel of personality captures more detailed aspects of psy-chological processes than previous work, and goes beyond

2 R. Sun, N. Wilson / Cognitive Systems Research 29–30 (2014) 1–30

specialized computational models addressing personalityalone (e.g., Shoda & Mischel, 1998). It is necessary to gobeyond abstract notions of goals, plans, cognitive affectiveunits, and so on—It is one thing to argue that personalitytraits consist of configurations of goals, plans, or cognitiveaffective units, it is quite another to map personality traitsto more concrete and better grounded psychologicalprocesses and mechanisms. It is useful to ground personal-ity traits in a cognitive architecture, so that they can beexplained in a deeper and more unified way, along withmany other psychological phenomena, based on the sameprimitives as envisaged within a generic cognitive architec-ture (without any significant additions or modifications).Also, coupled with the account of learning in the cognitivearchitecture, a model of personality should be able toaccount for the emergence, shaping, and tuning ofpersonality by a variety of factors.

Therefore, personality should be computationallycaptured and explained by a computational cognitivearchitecture (Sun, 2002, 2009). That is, it should be cap-tured and explained based on adequate built-in representa-tion of basic motivational, metacognitive, action decisionmaking, reasoning, and other processes, within a (rela-tively) generic and comprehensive computational modelof the mind. Such representations and related mechanismsand processes may capture the interaction of internally feltneeds and external environmental factors in determininggoals and actions by individuals. They may capture therelative invariance within an individual in terms of behav-ioral propensities and inclinations at different times and withregard to different situations (social or physical), as well asbehavioral variability. The (relative) invariance of personal-ity has been extensively argued for in personality psychology(e.g., Caprara & Cervone, 2000; Epstein, 1982; Hall &Gardner, 1985; Maddi, 1996; Ryckmann, 1993; Staub,1980; Wiggins, 1996); it can be captured computationally.

An outline of our general framework can be summa-rized as follows (cf. Cervone, 2004; Mayer, 2005; Readet al., 2010; Winter, John, Stewart, Klohnen, & Duncan,1998): Within a comprehensive cognitive architecture, thereis constant interaction among its subsystems: themotivational subsystem, the metacognitive subsystem, theaction-centered (procedural) subsystem, and the non-action-centered (declarative) subsystem. Within themotivational subsystem, there is a set of basic motives,termed primary drives, which are universal across individu-als. Individual differences may be explained (in part) by thedifferences across this set of drives in terms of drivestrengths (activations) in different situations by differentindividuals. These drives, with their different strengths, leadto setting of different goals as well as major cognitiveparameters by the metacognitive subsystem. Individualdifferences in terms of drive strengths are consequentlyreflected in the resulting goals as well as majorcognitive parameters. On the basis of the goals set andthe cognitive parameters chosen, an individual makesaction decisions, within the action-centered subsystem,

possibly in consultation with the non-action-centered sub-system using its declarative knowledge. Thus their actionsreflect their fundamental individual differences as well assituational factors as a result. Their actions in turn affectthe world in which they act.

As such, personality results from the complex interac-tion of a number of psychological entities, mechanisms,and processes, as well as their interaction with the world.Hence, computational simulations of personality basedon a cognitive architecture can be useful in supplementingalternative approaches: Computational modeling and sim-ulations may enable us to see exactly how these entities,mechanisms, and processes interact.

Why are all these mechanisms and processes needed toaccount for human personality? It is conceivable that theseindividual aspects or components of the mind exert a con-siderable influence individually or collectively on the per-sonality of an individual. For example, Sun and Wilson(2011) and Deci (1980) argued that motivational represen-tations are important for capturing personality. Similarly,behaviors (actions) are important, at least for measuringpersonality (Sun, 2002). Existing computational modelsof personality tend to be specialized, missing, for example,well-developed models of motivation, reasoning, metacog-nition, and so on. They tend to ignore some essential psy-chological distinctions (e.g., implicit vs. explicit; morelater). They often rely on extensive parameter tweakingto generate desired outcomes. The upshot is that a modelof personality should be a model of cognitive architecture(or a model of personhood, more generally), including allof these aspects enumerated earlier (and possibly more).

This is not to say that other aspects of personhood,besides a cognitive architecture, such as various aspectsof the biological or social being, are not important. Amodel of personality should ideally be a model of person-hood, biological, psychological, and social. The fact, how-ever, is that there is currently no comprehensive model ofthe physical and physiological person available, beside rel-atively comprehensive models of the mind, that is, compu-tational cognitive architectures. Given that we currentlyonly have cognitive architectures available, which presum-ably capture some of the most important parts of thepersonhood, it seems wise to start with cognitive architec-tures. (Sociocultural aspects are dealt with elsewhere; seeSun, 2006, 2012; Wilson & Sun, in preparation.)

Thus, a variety of computational simulations of person-ality can be performed through a cognitive architecture assketched above. The simulations should naturally fall outof the cognitive architecture, without any significant alter-ations or additions. They may help to clarify issues ofpersonality in an exact, mechanistic, and process-basedway. Practically speaking, the resulting personality modelmay be applied in practically useful ways. In particular, amodel of personality may be important for simulatingsome social phenomena; that is, it may be applied tocognitive social simulation (Sun, 2006, 2012), which isone of our goals for the future.

R. Sun, N. Wilson / Cognitive Systems Research 29–30 (2014) 1–30 3

In the remainder of the paper, first an abstract model ofpersonality, consisting of a number of general principles(derived from the CLARION cognitive architecture), ishypothesized and briefly justified. The CLARION cogni-tive architecture itself is then sketched, which providesthe basis for computational simulation. A series of simula-tions is then presented, which shows capabilities of themodel. A general discussion follows, which includescomparisons with existing computational models.

2. Principles of personality

Below, we present an abstract model of human person-ality in the form of a set of basic principles, before describ-ing a computational cognitive architecture, CLARION,from which these principles were derived, and before detail-ing computational simulations of personality withinCLARION.

A number of basic principles (as derived from CLAR-ION) may be enumerated and briefly justified as follows(see Sun & Wilson, 2011; Sun, Wilson, & Mathews, 2011;Wilson & Sun, in preparation; cf. Cervone, 2004; Mayer,2005; McAdams & Pals, 2006; Shoda & Mischel, 1998).

2.1. Principle 1

2.1.1. Principle 1: Embodying personality in a cognitive

architecture

Personality should emerge from the interaction amongvarious components of the mind. That is, mechanistically,it should emerge from the processes of, and informationflows among, various subsystems (and modules within) ofa computational cognitive architecture. The cognitivearchitecture should allow the emergence of differentpersonality types, and the adaptation of personalitythrough experience (to some extent).

2.1.2. Brief justificationThis is the fundamental principle on which the present

work is based. This principle has been argued in Section 1right from the beginning. Therefore, these arguments willnot be repeated here; see Section 1 for details. Variousaspects related to this principle, on the other hand, willbe articulated and justified next.

2.2. Principle 2

2.2.1. Principle 2: The dichotomy of implicit and explicit

psychological processes in the cognitive architecture

Psychological processes may be either implicit orexplicit; the two types co-exist but are different in severalimportant respects, with implicit processes often beingmore fundamental.

2.2.2. Brief justification

Although there have been disagreements concerningthe notion of conscious accessibility (and therefore the

implicit–explicit distinction), it is generally agreed uponthat at least some part of performance is not consciouslyaccessible under normal circumstances. Reber (1989), forexample, argued that “although it is misleading to arguethat implicitly acquired knowledge is completely uncon-scious, it is not misleading to argue that implicitly acquiredepistemic contents of mind are always richer and moresophisticated than what can be explicated” (p. 229).Experimental data supporting the implicit–explicit distinc-tion have been thoroughly reviewed by, for example, Reber(1989), Seger (1994), and Sun (2002).

Generally speaking, mechanistically, explicit processingmay be described as based on rules in some way, whereasimplicit processing is more associative (Sun, 2002). Explicitprocessing may involve the manipulation of symbols. Thenature of symbol manipulation allows systematicity insome way. In contrast, implicit processing involves moreinstantiated knowledge that is holistically associated, andtherefore implicit processing is often more situation-spe-cific and based on approximate match (Sun, 1994, 2002;Reber, 1989). Empirical evidence in support of thesedifferences can be found in Seger (1994), Cleeremans,Destrebecqz, and Boyer (1998), and Sun (2002).

Another distinction between explicit and implicit pro-cesses is in terms of attention. While explicit processes usu-ally require attention, implicit processes usually do not(Reber, 1989). This difference may be demonstrated, forexample, through a dual-task setting that cancels the effectof explicit knowledge without affecting implicit processes.Such attention differences have been found in serial reac-tion time tasks, artificial grammar learning tasks, dynamiccontrol tasks, and other tasks (Cleeremans et al., 1998;Curran & Keele, 1993; Sun, Slusarz, & Terry, 2005).

Implicit processes are often believed to be more funda-mental. The fundamental importance of implicit processeshas been argued for by Reber (1989; see the quote above),as well as many others (Sun, 2002; Sun et al., 2005).

2.3. Principle 3

2.3.1. Principles 3: The division of drives and goals in the

motivational (sub)system

Human motivation may be explained by a combinationof implicit drives and explicit goals, with drives being morefundamental.

Let us examine specifically the issue of explicit vs. impli-cit motivation within the motivational (sub)system. On theone hand, it is hard to imagine that there is no explicit rep-resentation of goals, since all the evidence points to thecontrary (e.g., from the skill learning literature). On theother hand, the internal processes of drives, needs, ordesires do not appear to be readily accessible (i.e., notexplicit; Hull, 1943; Murray, 1938; Sun, 2009; Winteret al., 1998). So, it is reasonable to assume that (1) thegeneral idea of dual representation (implicit and explicit)

as argued above (Principle 2) is applicable to motivationand (2) correspondingly, implicit motivational processesare more fundamental than explicit motivational processes.

Explicit motivational representation consists mainly ofexplicit goals (Anderson & Lebiere, 1998; Sun, 2009).Explicit goals provide specific, tangible motives for actions,often helping to mobilize resources. Explicit goal represen-tations may facilitate explicit cognitive processes to workon the achievement of these goals, in addition to involvingimplicit processes. Explicit goals also allow more behav-ioral flexibility and formation of expectancies (Epstein,1982). It may sometimes be necessary to compute a matchof a state of the world to the goal (so as to discern the pro-gress in achieving the goal, as well as to generate context-dependent reinforcement signals), which may be facilitatedby using an explicit representation of goals. Explicit goalrepresentations are more persistent, while implicit motivesmay be more likely to change from moment to moment.

However, the more fundamental part of the motiva-tional (sub)system, the implicit motives, consists of“drives“ (i.e., basic needs, basic desires, intrinsic motives,and so on; see Sun, 2009). A generalized notion of “drive”is adopted in the present work, different from stricter inter-pretations.1 In the present work, drives denote internallyfelt needs of all kinds that likely may lead to correspondingbehaviors, regardless of whether the needs are physiologi-cal or not, whether the needs may be reduced by the corre-sponding behaviors or not, or whether the needs are forend states or for processes (thus it transcends controversiessurrounding the stricter notions of drive). Drives may beactivated through stimuli from external and internalsources (modulated by internal sensitivity parameters).

Human motivation is known to be highly complex andvaried (Weiner, 1992), and cannot be captured with simpleexplicit goals alone. For example, the interactions of drivesrequire a more complex representation. Their changes overtime, which are often gradual, also require a more quanti-tative and graded representation. Furthermore, Maslow(1943) and Murray (1938) specifically argued for theunconscious characteristics of “needs” (see also Hull,1951). Given all of the above, it is reasonable to hypothe-size that implicit, graded processes of drives are necessary.

Empirical evidence from social psychology also pointsto the duality of human motivation. For example, Woodand Quinn (2005) explored the duality of motivations ineveryday life, and the relationship between implicit andexplicit motivations, in ways analogous to the analysis ofimplicit and explicit processes in general in Sun et al.

1 In the past, Hull (1951) developed the most detailed conception of“drives”—an implicit, pre-conceptual representation of motives. In hisview, drives arose from need states, behaviors were driven so as toeliminate need states, and drive reduction was the basis of reinforcement.Although Hull’s conception of drive had significant explanatory power, itfailed to capture many motivational phenomena—the variety of differentmotivations proved too difficult to be encompassed by his theory of drive.A more general notion is therefore needed.

(2005). Aarts and Hassin (2005) reviewed evidence of bothexplicit and implicit motivations in human behavior.

Implicit motivational processes are believed to be morefundamental, because on their basis, explicit goal represen-tations arise, which clarify and make concrete implicitmotivational dynamics (Sun, 2009). Castelfranchi (2001),for example, discussed such implicit-to-explicit motiva-tional processes, in ways analogous to general implicit-to-explicit cognitive “emergence” (Sun, 2002). See Principle4 in this regard.

2.4. Principle 4

2.4.1. Principle 4: Goal setting on the basis of drivesGoals may be determined on the basis of the interaction

and competition of different drives, as a result of situa-tional inputs and internal factors.

As argued before, both drives and goals are needed in amotivational (sub)system (Principle 3). However, a ques-tion is how goals should be determined. Because of the factthat drives represent fundamental needs, desires, andmotives, it is natural to determine goals on the basis ofdrives (for the most part). As argued by, for example, Tol-man (1932), Murray (1938), Simon (1967), Deci (1980), andWright and Sloman (1997), explicit goals may be set on thebasis of implicit, embodied motives (namely, drives).Explicit goals may be set (possibly stochastically) basedon the competition of drives activated from situational cuesand internal factors (Sun, 2009). The drives provide thecontext within which explicit goals are chosen.

Conceivably, motivational processes may be tuned tosome extent through experience (Murray, 1938; Toates,1986; Weiner, 1992). In particular, the selection of goalson the basis of drives is learnable through experience insociocultural and physical environments.

2.5. Principle 5

2.5.1. Principle 5: The fundamental role of motivation

(drives and goals) in determining personality

Personality is rooted in motivational processes, and inparticular in implicit drives.

2.5.2. Brief justificationFundamental behavioral traits, that is, personality,

should map onto essential motivational processes (at leastfor a significant part; Deci, 1980; Murray, 1938). Thus per-sonality should map onto drives, because drives are themost fundamental and the most stable (Principle 4). Otherprocesses may be more transient, due to environment,learning, and adaptation. Although drives may be tunedalso, they are, relatively speaking, more stable than otherprocesses. Therefore, it is reasonable to ground personality,first and foremost, in drives and then goals on their basis(Principle 3 and 4).

Using mostly motivational processes to account forhuman personality requires principled justifications, whichmay be done from a variety of perspectives: philosophy,religion, the psychology literature on personality, and theexisting computational models of personality.

First, Schopenhauer (1819) contends that the ultimateprinciple is “Will”—a non-rational urge at the foundationof being. Existence is, in essence, endless striving with blindimpulses, which are the forces within humans that haveprecedence over reason and rationality. In particular, heidentifies the Will to Live. Such non-rational drives definethe essential human condition. Similarly, in the Buddists’conception of the world, desire lies at the root of humanexistence. According to Buddhism, life is a never-endingflow of desire. Everyday life is a transient, impermanentsequence of circumstances driven by changing desires. Atan abstract level, these thoughts are similar to our viewof the fundamental role of motivation.

Existing work in social-personality psychology alsoshows how personality can be closely related to humanmotivation. Deci (1980) made an elaborate case for thispoint, reviewing the literature on motivation and personal-ity and arguing for their close relationship. Reiss (2010)argued that “Everybody is motivated by the . . .basicdesires, but people prioritize them differently. Every personhas his or her own hierarchy, which is highly correlated tonormal personality traits . . .A powerful predictor ofbehavior in natural environments [personality] is how aperson prioritizes the . . .basic desires.” Shoda and Mischel(1998) argued that personality could be understood interms of cognitive-affective units, for example, goals,plans, expectancies, and so on. Individual differences inpersonality might emerge on the basis of cognitive-affectiveunits.

Some existing computational models of personality relyon motivational representations. For example, Read et al.(2010) developed a computational model of personalitybased on “goals”. However, in our view (see Principle 4),goals need to be set based on underlying “basic needs” or“basic desires” (as argued by Deci, 1980; Murray, 1938;Sun, 2009; Tolman, 1932). Similarly, Shoda and Mischel(1998) proposed a computational model that was basedon cognitive-affective units (e.g., goals, plans, and expec-tancies). The same may be said regarding the PSI modelof Schaub (2001) and Doerner (2003). In sum, there are amyriad of reasons for basing personality on motivation.

It should be noted that personality involves more thanjust motivation (Principle 1). It may involve a variety ofpsychological mechanisms and processes, beyond motiva-tion, although they may be somewhat less important topersonality. Therefore, personality types, besides beingmapped onto motivational structures, representations,and processes, need also be mapped onto other mecha-nisms and processes. The determination of personalitytypes involves various motivational, metacognitive, cogni-tive, perceptual, and other parameters (Schaub, 2001;Sun & Wilson, 2011).

2.6. Principle 6

2.6.1. Principle 6: The role of action decision making in

personality

Action decision making (i.e., procedural processes) onthe basis of the goal chosen and the situational inputs isan integral part of personality.

Here action is defined in a broad sense, including bothphysical and mental actions (Sun, 2002, 2003). Actions(behaviors) are the ultimate measure of personality.Without it, there would be no objective way of classifyingpersonality types. Therefore, action decision making (pro-cedural processes) is important to measuring personality.Moreover, it is also an important part of personalitybecause it is integral to the underlying dynamics ofpersonality.

For example, consistent with Principle 1, the personalitytrait of being dominating may be captured by a drive statewhere dominating others is emphasized (Principle 5), a spe-cific goal being chosen, actions, routines, and plans forachieving that goal being carried out, and reasoning relatedto that goal being applied (Principle 7 later). In terms ofactions (behaviors), a dominating person, when in relevantsituations and reacting to related cues, exhibits dominatingbehaviors regularly. On the basis of such behaviors(actions), the individual is viewed as a dominating person.

This process produces dominating behaviors across anarray of situations, only if other drives and goals leadingto competing behaviors are not as frequently and asstrongly activated. Another prerequisite is the learned con-nections between the goal and the proper actions within agiven sociocultural and physical environment. Proceduralknowledge and skills (necessary for action selection) arelearnable, as commonly accepted (Anderson & Lebiere,1998; Montague, 1999; Sun, 2002), and they are subjectto sociocultural influences. Actions, therefore, form a partof the chain that leads to personality traits.

2.7. Principle 7

2.7.1. Principle 7: The role of declarative knowledge and

reasoning in personality

Declarative knowledge and reasoning affect personalitythrough affecting actions (behaviors), although their effectsare less direct.

Cervone (2004) argued for the importance of knowl-edge, belief, schema, appraisal, reasoning, and so on, thatis, declarative knowledge and processes, as determinantsof personality.

As an example, the personality trait of being dominatingmay be captured by a drive state where dominating othersis emphasized, a goal of dominating others being chosen,actions for achieving that goal being carried out, and so

on. But beyond these, reasoning related to that goal mayalso be carried out, for example, regarding whether one’sactions would actually achieve the goal. Such reasoning,as well as the declarative knowledge on which the reason-ing is based, is important to behavior choices and thereforeto personality.

However, reasoning and declarative knowledge canimpact personality only on the basis of drives and goals,and affect procedural processes (action decision making)only indirectly. Declarative knowledge and processes arelearnable and flexible, subject to sociocultural influences.

3. A cognitive architecture as a model of personality

Based on the discussions so far, desirable attributes of anew model of personality may include:

� The new model should result directly from a relativelycomprehensive computational framework of the mind,that is, a computational cognitive architecture (asargued earlier), thus situating personality in a broadcontext;� The new model should have well-developed and pre-

cisely-specified mechanistic and process-based (i.e., com-putational) details;� In particular, the new model should be based on a well-

developed model of essential motives of human behav-ior, thus grounding (to a significant extent) personalityin a general theory of human motivation (Sun, 2009);� The new model should make direct contact with empir-

ical data and be able to capture and explain such data.

To address the first three attributes, we will look intoCLARION below. The last one will be addressed in thenext section on the basis of CLARION.

3.1. Overview of CLARION

CLARION is a generic cognitive architecture—arelatively comprehensive model of psychological processesof a wide variety, specified computationally. It has beendescribed in detail and justified on the basis of psycholog-ical data in Sun (2002, 2003; see also Sun, Merrill, & Peter-son, 2001; Sun et al., 2005 and Helie & Sun, 2010). It can bethe framework within which personality emerges. That is,CLARION may serve as a model of personality as is, with-out having to build a specialized model of personality.

CLARION consists of a number of subsystems: theaction-centered subsystem (the ACS), the non-action-centered subsystem (the NACS), the motivational subsys-tem (the MS), and the metacognitive subsystem (theMCS). The role of the action-centered subsystem is tocontrol actions (regardless of whether the actions are forexternal physical movements or for internal mental opera-tions), utilizing and maintaining procedural knowledge.The role of the non-action-centered subsystem is tomaintain and utilize declarative knowledge. The role of

the motivational subsystem is to provide underlying moti-vations for perception, action, and cognition (in terms ofproviding impetus and feedback). The role of the metacog-nitive subsystem is to monitor, direct, and modify the oper-ations of the other subsystems dynamically.

Each of these interacting subsystems consists of two“levels” of representations (i.e., a dual-representationalstructure) as theoretically posited in Sun (2002). Generallyspeaking, in each subsystem, the “top level” encodes explicit

knowledge (using symbolic/localist representations) and the“bottom level” encodes implicit knowledge (using distrib-uted representations; Rumelhart, McClelland, & the PDPResearch Group, 1986). The two levels interact, for exam-ple, by cooperating in action decision making, through inte-gration of the action recommendations from the two levelsof the ACS respectively, as well as by cooperating inlearning through a “bottom-up” and a “top-down” learningprocess (Sun et al., 2001, 2005). See Fig. 1 for a sketch.

As has been pointed out before, existing theories tend toconfuse implicit (reflexive) and explicit (deliberative)processes. Hence the “perplexing complexity” (Smillie,Pickering, & Jackson, 2006). In contrast, CLARION gener-ally separates and integrates implicit and explicit processes ineach of its subsystems. With such a framework, CLARIONcan provide better explanations of empirical findings in awide range of domains (see, e.g., Helie & Sun, 2010; Sun& Wilson, 2011; Sun et al., 2001, 2005 for details).

Another particularly important characteristic of this cog-nitive architecture is its focus on the cognition–motivation–environment interaction, as opposed to dealing only withcognition in the narrow sense.

Below, we will examine each of the subsystems in moredetail (which will illustrate some of these points).

3.2. The action-centered subsystem

The action-centered subsystem (the ACS) captures theaction decision making of an individual when interactingwith the world.

In the ACS, the process for action selection is essentiallythe following: Observing the current (observable) state ofthe world, the two levels within the ACS (implicit or expli-cit) make their separate action decisions in accordance withtheir respective procedural knowledge, and their outcomesare “integrated”. Thus, a final selection of an action ismade and the action is then performed. The action changesthe world in some way. Comparing the changed state of theworld with the previous state somehow, the person learns.The cycle then repeats itself. Thus, the overall algorithmfor action selection is as follows:

1. Observe the current input state x (including the currentgoal).

2. Compute in the bottom level the “value” of each of thepossible actions (ai’s) associated with the current state x:Q(x,a1),Q(x,a2), . . . ,Q(x,an). Stochastically choose oneaction according to these values.

Fig. 1. The CLARION cognitive architecture.

3. Find out all the possible actions at the top level (b1,b2,. . . ,bm), based on the current state x (which goes upfrom the bottom level) and the existing rules in placeat the top level. Stochastically choose one action.

4. Choose an action, by stochastically selecting theoutcome of either the top level or the bottom level.

5. Perform the action, and observe the next input state y

and (possibly) the reinforcement r.6. Update knowledge at the bottom level in accordance

with an appropriate learning algorithm (e.g., Q-learning;more later), based on the feedback information.

7. Update the top level using an appropriate learningalgorithm (e.g., the RER algorithm; more later).

8. Go back to Step 1.

In this subsystem, the bottom level is implemented withneural networks involving distributed representations(Rumelhart et al., 1986), and the top level is implementedusing symbolic/localist representations.

The input state (x) to the bottom level consists of thesensory input (environmental or internal), the current goal,and so on. All that information is important in deciding onan action. The input state is represented as a set of dimen-sional values (microfeatures): (d1,v1)(d2,v2)....(dn,vn). Theoutput of the bottom level is the action choice, also repre-sented as a set of dimensional values.

At the top level, “chunk” nodes are used for denotingconcepts. A chunk node connects to its correspondingdimensional values (microfeatures) represented as separate

nodes in the bottom level (constituting a distributed repre-sentation in the bottom level). At the top level, action rulesconnect chunk nodes representing conditions to chunknodes representing actions.

At the bottom level of the ACS, using neural networksencoding implicit knowledge, actions are selected basedon their Q values. A Q value is an evaluation of the “qual-ity” of an action in a given input state: Q(x,a) indicateshow desirable action a is in state x. At each step, given statex, the Q values of all the actions (i.e., Q(x,a) for all a’s) arecomputed in parallel. Then the Q values are used to decidestochastically on an action to be performed, through aBoltzmann distribution of Q values:

pðajxÞ ¼ eQðx;aÞ=s=X

eQðx;aiÞ=s

where s (temperature) controls the degree of randomness ofaction decision making, and i ranges over all possible ac-tions. (This is known as Luce’s choice axiom; Watkins,1989.)

For learning implicit knowledge at the bottom level (i.e.,the Q values), the Q-learning algorithm (Watkins, 1989), areinforcement learning algorithm, may be used. Q valuesare gradually tuned through successive updating, whichenables reactive sequential behavior to emerge throughtrial-and-error interaction with the world (Sun et al.,2001; Watkins, 1989).

For learning explicit action rules at the top level with a“bottom-up” learning process (Karmiloff-Smith, 1986;

Fig. 2. The structure of the motivational subsystem.

2 Note that a generalized notion of “drive” is adopted in CLARION,different from the stricter interpretations of drives Hull (1951). Asdiscussed in Sun (2009), it is a generalized notion that transcendscontroversies surrounding the stricter notions of drive.

Sun et al., 2001), the Rule-Extraction-Refinement algorithm(RER) relies on information from the bottom level. Thebasic idea is as follows: If an action chosen (by the bottomlevel of the ACS) is successful (i.e., it satisfies a certain cri-terion), then an explicit rule is extracted at the top level.Then, in subsequent interactions with the world, the ruleis refined by considering the outcome of applying the rule:If the outcome is successful, the condition of the rule maybe generalized to make it more universal; if the outcome isnot successful, then the condition of the rule should bemade more specific (Sun et al., 2001).

On the other hand, top-down learning goes in theopposite direction. With explicit knowledge at the top level,the bottom level learns under its guidance. That is, initially,the individual may rely mostly on the action rules at the toplevel for its action decision making. But gradually, whenmore and more knowledge is acquired by the bottom levelthrough “observing” actions directed by the top-level rules(based on the same reinforcement learning mechanism asdescribed before), the individual relies more and more onthe bottom level. Hence, top-down learning takes place(Sun, 2002).

For stochastic selection of the outcomes of the twolevels, at each step, with probability PBL, the outcome ofthe bottom level is used. Likewise, with probability PRER,if there is at least one RER rule indicating an action inthe current input state, the outcome from that rule set(determined through competition) is used. Other compo-nents, if they exist, may also be included in the stochasticselection. There exists some psychological evidence forsuch intermittent use of rules (Sun et al., 2001). The selec-tion probabilities may be variable, determined through aprocess of “probability matching”: the probability ofselecting a component is determined based on the relativesuccess ratio of that component.

3.3. The non-action-centered subsystem

The non-action-centered subsystem (the NACS) is fordealing with declarative knowledge, which is not action-centered. It stores such knowledge in a dual representa-tional form (the same as in the ACS): that is, in the formof explicit “associative rules” (at the top level), and in theform of implicit “associative memory” (at the bottomlevel). Its operation is under the control of the ACS.

At the bottom level of the NACS, associative memorynetworks encode implicit non-action-centered (declarative)knowledge. Associations are formed by mapping an inputto an output (e.g., the Backpropagation learning algorithmmay be used to establish such associations; Rumelhart et al.,1986.)

At the top level of the NACS, explicit non-action-centered (declarative) knowledge is stored. As in theACS, “chunk” nodes (denoting concepts) at the toplevel are linked to dimensional values (microfeatures)represented at the bottom level. Additionally, in the toplevel, links between chunk nodes encode explicit associative

rules (explicit associative rules may be learned in a varietyof ways; Sun, 2003).

As in the ACS, top-down or bottom-up learning maytake place in the NACS, either to extract explicit knowledgeat the top level from the implicit knowledge in the bottomlevel, or to assimilate the explicit knowledge of the top levelinto the implicit knowledge in the bottom level.

With the interaction of the two levels, the NACS carriesout rule-based, similarity-based, and constraint-satisfaction-based reasoning. These mechanisms will notbe covered here because they are not utilized in this paper(details can be found in, e.g., Helie & Sun, 2010).

3.4. The motivational subsystem

The motivational subsystem (the MS) is concerned withwhy an individual does what he/she does. The relevance ofthe MS to the ACS lies primarily in the fact that it providesthe context in which goals and reinforcements of the ACSare determined. It thereby influences the working of theACS (and by extension, the working of the NACS).

A dual motivational representation is in place. Theexplicit goals (such as “finding food”), which are essentialto the working of the ACS as explained before, may be gen-erated based on implicit drives (e.g., “being hungry”). Theexplicit goals derive from, and hinge upon, implicit drives.See Fig. 2 for a sketch of the motivational subsystem. Forjustifications, see the principles in Section 2 (see also Deci,1980; Sun, 2009; Wright & Sloman, 1997).

3.4.1. Primary drives

“Primary drives” are essential to an individual and aremost likely built-in (hard-wired) to a significant extent tobegin with. Some sample low-level primary drives include:food, water, reproduction, and so on. Beyond such low-levelprimary drives concerning mostly physiological needs,there are also high-level primary drives: for example, dom-inance and power, fairness, and so on (see McDougall, 1936;Murray, 1938; Reiss, 2010; Sun, 2009; see also Toates,1986; Weiner, 1992).2

Table 1Primary drives.

Food The drive to consume nourishmentWater The drive to consume fluidSleep The drive to rest and/or sleepReproduction The drive to mateAvoiding Danger The drive to avoid situations that have the potential to be or already are harmfulAvoiding Unpleasant Stimuli The drive to avoid situations that are physically (or emotionally) uncomfortable or negative in natureAffiliation & Belongingness The drive to associate with other individuals and to be part of social groupsDominance & Power The drive to have power over other individuals or groupsRecognition & Achievement The drive to excel and be viewed as competent at somethingAutonomy The drive to resist control or influence by othersDeference The drive to willingly follow and serve a person of a higher status of some kindSimilance The drive to identify with other individuals, to imitate others, and to go along with their actionsFairness The drive to ensure that one treats others fairly and is treated fairly by othersHonor The drive to follow social norm and code of behavior and to avoid blamesNurturance The drive to care for, or to attend to the needs of, others who are in needConservation The drive to conserve, to preserve, to organize, or to structure (e.g., one’s environment)Curiosity The drive to explore, to discover, and to gain new knowledge

The primary drives (low-level and high-level) may beexplained roughly as in Table 1. This set of primary driveshas been extensively explained and justified in priorwritings (e.g., Sun, 2003, 2009; Sun & Wilson, 2011; seealso the principles in Section 2), based on work in socialpsychology as well as early work of ethology.3

3.4.2. Details of drive processing

The internal processing of the drives involves the follow-ing modules:

(1) The drive initialization module: It carries out (a) amapping from a personality type to the initial“deficits” of the drives, and (b) a mapping from a per-sonality type to the “baselines” of the drives. Thesetwo mappings are relatively fixed (that is, they areoften performed only once for an entire simulation).See the justifications in relation to personality earlier.

(2) The drive preprocessing module: For each drive d,there is a “preprocessor” (a neural network) thatpicks out relevant information for determining thedrive-specific stimulus level (which is an evaluationof the relevance of the current situation with regardto the drive, used for calculating the drive strengthlater). It is a built-in detector for relevant informationin relation to drive d.4

(3) The drive core module: It determines drive strengths(using a neural network) based on:

dsd ¼ gainu � gains � gaind � stimulusd � deficitd

þ baselined

3 Briefly, this set of hypothesized primary drives bears close relationshipsto Murray’s needs (1938), Reiss’s motives (2010), Schwartz’s (1994) 10universal values, and so on. The prior justifications of these frameworksmay be applied, to a significant extent, to this set of drives as well (seeMaslow, 1943; McDougall, 1936; Murray, 1938; Reiss, 2010; Sun, 2009).

4 This mapping may include generalizations from some familiar scenar-ios to other scenarios (accomplished by neural networks; Sun, 2003).

where dsd is the strength of drive d, gaind is the individ-ual gain for drive d, gainu is the universal gain affectingall drives, gains is the gain affecting all the drives of onetype (e.g., the approach or the avoidance type; Elliot &Thrash, 2002), stimulusd is a value representing howpertinent the current situation is to drive d, deficitd indi-cates the perceived deficit in relation to drive d (whichrepresents an individual’s intrinsic sensitivity and incli-nation toward activating drive d), and baselined is thebaseline strength of drive d.

(4) The deficit change module: This module (a neuralnetwork) determines how deficitd changes over time,for each state-action step, in relation to the inputstates encountered, the goals adopted, the actionsperformed, and so on.

The justifications for the mappings may be found in avariety of literatures, ranging from ethological researchand modeling (e.g., Toates, 1986; Tyrell, 1993) to cognitivemodeling (e.g., Doerner, 2003; Schaub, 2001; Sun, 2009).In particular, the multiplicative combination of stimulusd

and deficitd has been argued for (e.g., Sun, 2009; Tyrell,1993). Note that these modules above are all Backpropaga-tion neural networks.

3.5. The metacognitive subsystem

The existence of the large number of drives and the goalsresulting from them leads to the need for metacognitivecontrol and regulation. Metacognition refers to one’sknowledge (implicit or explicit) concerning one’s own cog-nitive processes. Metacognition includes active monitoringand consequent regulation and orchestration of theseprocesses (Flavell, 1976; Reder, 1996). In CLARION, themetacognitive subsystem (the MCS) is closely tied to themotivational subsystem (the MS). The MCS monitors,controls, and regulates cognitive processes (Simon, 1967;Wright & Sloman, 1997). Control and regulation may be

in the forms of setting goals (which are then used by theACS) on the basis of drives, generating reinforcement sig-nals for the ACS for learning (on the basis of drives andgoals), interrupting and changing on-going processes inthe ACS and the NACS, setting essential parameters ofthe ACS and the NACS, and so on.

Structurally, this subsystem may be divided into a num-ber of functional modules, including:

� the goal module,� the reinforcement module,� the processing mode module,� the input filtering module,� the output filtering module,� the parameter setting module (for setting learning rates,

temperatures, etc.),

and so on.Let us look into one module. The goal module, in order

to select a new goal, first determines goal strengths for someor all of the goals, based on information from the MS (e.g.,the drive strengths and the current goal) and the current sen-sory input. This may be implemented through a Backprop-agation neural network for calculating goal strengths. Then,a new goal is stochastically selected on the basis of the goalstrengths. See the arguments earlier in support of goal set-ting on the basis of implicit motives (i.e., drives) (e.g., Deci,1980; Tolman, 1932; see also empirical work such as Over &Carpenter, 2009 and Elliot & Thrash, 2002).

In the simplest case, the following calculation may beperformed:

gsg ¼Xn

relevanced;s!g � dsd

where gsg is the strength of goal g, relevanced,s!g is a mea-sure of how relevant drive d is to goal g with regard to thecurrent situation s (which represents the support that drived provides to goal g), and dsd is the strength of drive d asgenerated by the MS. Once calculated, the goal strengthsare turned into a Boltzmann distribution and the new goalis chosen stochastically from that distribution.5

3.6. Simulations using CLARION

CLARION has been successful in simulating, account-ing for, and explaining a wide variety of psychologicaldata. For example, a number of well-known skill learningtasks have been simulated and explained using CLARIONthat span the spectrum ranging from simple reactive skillsto complex cognitive skills. Among them, serial reactiontime and dynamic control tasks are typical implicit learning

5 A goal persistence factor may be used to encourage persistence ofgoals. For example, a multiplicative goal persistence factor gpg P 1 maybe applied, if g is the current goal. The same goes for behavior persistencefactor within the ACS (bpb P 1).

tasks (mainly involving implicit processes), while Tower ofHanoi and alphabetic arithmetic are high-level cognitiveskill acquisition tasks (with significant presence of explicitprocesses). In addition, extensive work has been done mod-eling a complex, realistic minefield navigation task, whichinvolves complex sequential decision making.

Simulations have also been done with reasoning tasks,social simulation tasks, as well as metacognitive and moti-vational tasks. While accounting for various psychologicaldata, CLARION provides explanations that shed newlight on underlying psychological processes (see, e.g., Helie& Sun, 2010; Sun, 2002; Sun et al., 2001, 2005, 2011).

3.7. Personality based on CLARION

CLARION, by itself, can serve as a model of personal-ity, without any significant additions or modifications. Infact, it is a very generic model of personality. This is in con-trast to existing computational personality models, whichare highly specialized and deal only with capturing someaspects of personality and nothing else.

Let us relate the CLARION cognitive architecture topersonality. In CLARION, action decisions are made bythe ACS, but the action decisions are based on the currentgoal, which is (mostly) set by the MCS based on the drivesin the MS, possibly involving also reasoning within theNACS. Therefore, drive activation is the foundation ofbehavior, according to CLARION. The actions of an indi-vidual are directed by the flow of “desires” (i.e., drives)during the interaction with the world, that is, variousimpulses on a moment-to-moment basis. This is consistentwith our earlier argument that personality traits are, to alarge extent, motivationally-centered, so that personalityreflects largely the dynamics of the underlying motivationalsubsystem (Sun & Wilson, 2011).

Moreover, the CLARION personality model involvesdrive “deficits”, because within CLARION, given thecurrent input state (hence given the drive-specific stimuluslevels as described earlier), the drive strengths are primarilydetermined by drive deficits (along with gains and baselinesas discussed before). On that basis, goal setting, actionselection, and reasoning (e.g., in the service of action selec-tion) take place. Thus, within the CLARION framework,personality may involve a variety of parameters withinthe ACS, the NACS, the MCS, and the MS, with the MSbeing the most important part. A general outline of ourCLARION personality model is thus as articulated earlierin Introduction.

4. Personality simulations with CLARION

Our goal for the simulations is to show that CLARION,as a personality model, (1) can capture whatever existingcomputational personality models can capture, plus more,and (2) can match and explain existing human data relatedto personality. Therefore, in our simulations, we will use

existing simulation tasks as benchmarks of some sort, andmatch existing simulations as much as matching humandata.

To demonstrate that CLARION can capture what exist-ing specialized computational models of personality cancapture, we will show how CLARION captures the simula-tions of Read et al. (2010) and Quek and Moskowitz(2007); the latter work also provides quantitative matchingwith human data. Due to length considerations, not all pre-vious simulations by these and other authors can be fullydiscussed—The reader is referred to Sun and Wilson(2011), Sun et al. (2011), and Wilson and Sun (in prepara-tion) for additional simulations.

Beyond capturing what existing specialized computa-tional personality models capture, we will also discussbriefly a simulation demonstrating how parameter tweak-ing can be minimized in modeling personality, which alsoprovides matching of human data (through undergoinghuman-like personality disorder tests). This simulation willshow how CLARION captures personality disorders (with-out parameter tweaking).

4.1. Simulation 1

4.1.1. Simulation setup

This simulation was designed to show that the CLAR-ION personality model could respond to different situationswith behaviors appropriate for different personality types,and at the same time show sufficient behavioral variability.

Fifteen scenarios as shown in Table 2 (as in Read et al.,2010) were used to test the model. These scenarios wererepresented using a set of situational features as specifiedin Table 2 (as in Read et al., 2010).

Each scenario was presented to the CLARION personal-ity model for 100 time steps and the chosen behavior by themodel at each step was recorded. The process was repeatedfor 100 times for representing 100 different simulated “sub-jects”, each time with different random initial weights inneural networks to correspond to individual differences.

Let us look into the setup within CLARION. Within theMS, a neural network was pre-trained (using standardBackpropagation; learning rate = .01, momentum = 0) tooutput drive strengths as specified in the equation discussedearlier. deficitd was set as shown in Table 3 (for the generictype).6 baselined was set to a linear function of the initialdeficit: baselined = .1 * deficitd.7 stimulusd was as shown inTable 4. All gain parameters in the MS were set to 1.

The drive strengths from the MS were provided to theMCS. The goals that can be selected were as shown in

6 This subsection involves only a generic type, while a number of otherpersonality types will be tested in subsequent subsections. As discussedearlier, drive deficits, to a large extent, determine personality.

7 The deficit of a drive is “decayed” at each step over time using amultiplicative factor (decayd = 10%), when the simulated “subject” isactively addressing the drive (i.e., when the goal corresponds to the drive).This is a simplification for the sake of this simulation.

Table 5. A pre-trained neural network in the MCS (learn-ing rate = .01, momentum = 0) calculated goal strengths(as specified by the equation earlier; relevanced,s!g was asshown in Table 6). The goal strengths were then turnedinto a Boltzmann distribution (temperature = .01) and agoal was chosen stochastically from the distribution.

The chosen goal was provided to the ACS, which alsoreceived the current sensory inputs. The bottom level ofthe ACS with a Backpropagation neural network waspre-trained (learning rate = .01, momentum = 0) to deter-mine Q values. Rule Extraction and Refinement (RER)was involved at the top level during the process (extractionthreshold = .8), and extracted rules were used in the toplevel of the ACS (PRER = .2). The outputs by the ACS wereturned into a Boltzmann distribution (temperature = .05)and one action was stochastically chosen from that distri-bution (Table 7).8

4.1.2. Simulation results

The percentage of appropriate behaviors for these sce-narios was recorded, for the one most frequently chosenbehavior and for the top three most frequently chosenbehaviors, respectively. A behavior was considered appro-priate if its Q value as determined by the ACS was abovea certain threshold (.5) given the current scenario and themost plausible goal. The most plausible goal (the goal thatwas the most likely to be chosen given the scenario) was cal-culated based on the generic personality (see Table 3), whichallowed us to determine the appropriateness of behaviorswithout having any access to the actual goals being choseninternally by the simulated “subjects”; see Table 8.

The results, averaged over all 100 simulated “subjects”,can be seen in Table 9. In both measures (top 1 and top 3),behavior choices were substantially more accurate thanchance (�20%). The results provided some evidence forthe appropriateness of the model, in terms of both accuracyand variability, in the same way as was done by existingspecialized computational personality models (such asRead et al., 2010).

In relation to personality traits, some argued thatwithin-person variability was at least as high as between-person variability (Caprara & Cervone, 2000). Such vari-ability is consistent with CLARION. As explained earlier,in CLARION, the activations of drives and the selection ofgoals and actions are dependent on input stimuli from sit-uations. As different situations are encountered, differentdrives may be activated, and the activated drives“compete” with each other for the control of behaviorsthrough setting goals. Depending on what drives aresimultaneously activated, the goal is stochastically deter-mined by the competing drives (in part determined by thesituational input). Furthermore, once the goal is set, differ-ent behaviors (actions) compete with each other to bechosen through stochastic selection, on the basis of

8 A multiplicative behavior persistence factor was applied (bpb = 1.5).

Table 2The 15 scenarios with their respective features.

Individual Assignment:

At work; in office; w/ 0 others; work to do; urgentWorking with one other:

At work, in office/at desk, work to do, urgent work, w/ 1 otherWorking together on urgent project:

At work; conference room; in an office; conflict situation; work to do; urgent; w/ 2 or more others; w/ subordinates; w/ disliked acquaintanceAt a group meeting:

At work; in an office; conference room; conflict situation; work to do; w/ 2 or more others; w/ friends; w/ boss; w/ disliked acquaintanceReview with Boss:

At work; w/ boss; office; conflict situation; urgentTaking a break with coworkers:

In break room; at work; TV; work to do; w/ 2 or more others; w/ friendsTaking a break by yourself:

In break room; at work; in office/at desk; work to do; w/ 0 othersParty at work:

Party, conference room, work, alcohol, work to do, w/ 2 or more others, w/ friends, w/ boss, w/ subordinates, difference > 7 yearsSocial Engagement At Boss’ House:

w/ boss; w/ strangers; w/ disliked acquaintance; w/ friends; w/ 2 or more others; conflict situation; w/ subordinates; w/ romantic partnerDance:

Dancing; w/ friends; w/ potential date; w/ strangers; w/ 2 or more others; alcoholTrying to get a date:

party; restaurant; alcohol; w/ 1 other; w/ potential dateOn a Date:

restaurant; alcohol; w/ 1 other; w/ dateFamily birthday party:

home; party; w/ 2 or more others; w/ romantic partner; w/ relatives; w/ kids; age differences > 7Wedding party at a fancy restaurant:

party; wedding/formal party; restaurant; dancing; alcohol; w/ 2 or more others; w/ friends; w/ romantic partner; w/ kids; age differences > 7Party in a restaurant that has a bar:

party; bar; restaurant; dancing; alcohol; w/ 2 others; w/ friends; w/ strangers; w/ potential date

Table 3Drive deficits corresponding to personality types.

Drives Sociable Shy Confident Anxious Responsible Lazy Generic

Food 0 0 0 0 0 0 .1Water 0 0 0 0 0 0 .1Sleep 0 0 0 0 0 0.5 .05Avoiding Danger 0 0.2 0 0.6 0 0.3 .2Reproduction 0.2 0 0.3 0 0 0 .1Avoiding the Unpleasant 0 0.6 0 0.7 0 0.7 .2Affiliation & Belongingness 0.9 0.2 0.3 0.6 0.2 0 .6Recognition & Achievement 0.5 0 0.8 0 0.8 0 .2Dominance & Power 0 0 0.7 0 0.2 0 .2Autonomy 0 0.3 0.6 0 0.7 0.2 .5Deference 0 0.7 0 0.8 0 0.3 .3Similance 0.5 0.8 0 0.8 0.1 0.8 .7Fairness 0.2 0 0 0.3 0.5 0 .1Honor 0.5 0 0.3 0.2 0.8 0 .5Nurturance 0.6 0 0 0 0.3 0 .4Conservation 0 0.4 0 0.6 0.7 0.1 .3Curiosity 0.6 0 0.5 0 0 0 .4

situational inputs. The multiple stochastic competitionsresult in varying behaviors both across situations and overtime.

We also wanted to see how sensitive the model was tonoise in the features of the scenarios. So the 15 differentscenarios were again tested over 100 steps and for 100

simulated “subjects”. However, noise was added to thescenarios by flipping the values of two of the input nodes(randomly chosen) for each scenario. The chosen behaviorat each step was recorded and the appropriatenessmeasures were calculated. The results were as shown inTable 10. The appropriateness of behavior choices

Table 4The drive stimulus parameter measuring how pertinent a scenario is to a drive. Scenarios determine the drive-specific stimulus levels.

Scenario Food Water Sleep Danger Reprod. Avoid A&B R&A D&P Aut. Def. Sim. Fair. Honor Nut. Cons. Cur.

1 0 0 0.15 0 0 0.15 0 0.6 0 0.3 0 0 0 0.15 0 0.15 0.152 0 0 0 0 0 0.3 0.6 0.9 0.3 0.15 0.3 0.3 0.15 0.3 0 0.15 0.153 0 0 0 0 0 0.3 0.6 0.9 0.5 0 0.3 0.3 0.5 0.3 0 0.15 04 0 0 0 0 0 0.45 0.9 0.9 0.6 0.15 0.6 0.8 0.6 0.3 0 0 0.155 0 0 0 0.3 0 0.3 0 0.6 0.15 0.3 0.95 0 0.6 0.3 0 0 06 0.6 0.6 0.3 0 0 0.6 0.6 0.6 0.15 0.3 0.15 0.7 0.15 0.15 0.15 0 0.37 0.6 0.6 0.6 0 0 0.8 0 0 0 0.6 0 0 0 0 0 0.15 0.38 0.6 0.8 0 0 0.3 0.7 0.9 0.6 0.15 0.15 0.3 0.7 0.15 0.15 0.15 0 0.39 0.3 0.6 0 0.15 0.3 0.7 0.9 0.7 0 0.15 0.7 0.7 0 0.3 0.15 0 0.310 0.15 0.6 0 0.15 0.7 0 0.8 0.8 0.15 0.6 0 0.7 0.15 0.4 0.4 0 0.311 0.3 0.7 0 0 0.95 0 0.3 0.3 0.15 0.6 0 0.15 0 0.3 0.7 0 0.312 0.6 0.6 0 0 0.99 0 0.7 0.3 0.15 0.4 0 0.3 0.15 0.3 0.8 0 0.413 0.7 0.7 0.15 0 0 0.15 0.7 0.6 0.5 0.15 0.3 0.6 0.3 0.15 0.95 0.4 0.314 0.9 0.95 0 0 0.7 0.15 0.8 0.6 0.15 0.3 0.3 0.6 0.4 0.3 0.7 0.15 0.715 0.9 0.95 0 0.15 0.8 0.6 0.8 0.3 0.15 0.4 0 0.7 0.15 0.3 0.15 0 0.4

Table 5Goals determined by the MCS.

Eat Be selfDrink FollowRest MimicFlee Be fairSex Follow codeAvoid Be caringFit in OrganizeStand out ExploreLead

9 The labels for these pairs were not precise. We used them herenevertheless, to correspond to Read et al. (2010).

(averaged over 100 simulated “subjects”) was well abovechance for both measures. The results showed the robust-ness of the model (in terms of accuracy and variability),in the same way as was done with existing specializedcomputational personality models (such as Read et al.,2010).

4.2. Simulation 2

4.2.1. Simulation setup

The possibility of different personalities within the modelshould be explored, now that initial validity of the modelwas established (Sun & Wilson, 2011). In addition, per-son–situation interaction also needed to be explored,because many past debates highlighted the importance ofperson–situation interaction (Caprara & Cervone, 2000).

In order to show that CLARION could capture whatexisting specialized computational personality models cancapture, in this simulation, we did what Read et al.(2010) did for demonstrating personality types. Six differ-ent personalities were set up as shown in Table 3 (as inRead et al., 2010). These personalities were designed toform three complimentary pairs: Sociable-Shy, Confident-Anxious, and Responsible-Lazy. Each of these pairs corre-sponded to the far ends of one of the three dimensions ofthe Five-Factor Model (John & Srivastava, 1999; Wiggins,

1996).9 One pair consists of the shy and the sociable, at thetwo ends of the extroversion dimension. Another pair con-sists of the anxious and the confident, at the two ends of theneuroticism dimension. The third pair consists of the lazyand the responsible, at the two ends of the conscientious-ness dimension. Their corresponding drive deficit levels,which determined these different personality types to a sig-nificant extent (as argued before), were as shown in Table 3.

This simulation used the same parameter settings as theprevious simulation. The model (for each of the six person-ality types) was run on the set of 15 scenarios. Each sce-nario was tested for 100 steps and the chosen behaviorswere recorded. The process was repeated for 100 differentruns (with different initial weights; representing 100 differ-ent simulated “subjects”).

Note that CLARION addressed individual differences inpersonality differently from Read et al. (2010). We manip-ulated drive deficit variables within the motivational sub-system to capture the differences, whereas Read et al.relied on altering specific goals and resources (as well asadditional “inhibition”). The present simulation exploredhow individual differences in drive deficits could translateinto behaviors indicative of personality types; in otherwords, it explained the observed tendencies of specific per-sonalities largely through drive deficits.

4.2.2. Simulation results

Figs. 3–5 show the results of the simulation with CLAR-ION using the three pairs of personality types (Sun &Wilson, 2011).

As shown by Figs. 3–5, the two personality types in eachpair behaved differently across this set of 15 scenarios.These figures were separated by personality pairs with thescenarios on the x-axis and the index of the most frequently

Table 6The relevance of drives to goals (for the sake of selecting a goal to pursue). The columns indicate drives and the rows indicate goals.

Goal Food Water Sleep Danger Reprod. Avoid A&B R&A D&P Aut. Def. Sim. Fair. Honor Nut. Cons. Cur.

Eat 0.95 0.05 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0Drink 0.05 0.95 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0Rest 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0Flee 0 0 0 0.95 0 0.1 0.05 0 0 0 0.05 0.05 0 0 0 0 0Sex 0 0 0 0 0.95 0 0 0.05 0.05 0 0 0 0 0 0.1 0 0Avoid 0 0 0.05 0.1 0 0.95 0.05 0 0 0 0.05 0.05 0.05 0 0 0 0Fit in 0 0 0 0 0.05 0 0.95 0 0 0 0.05 0.05 0.05 0.1 0.05 0 0Stand out 0 0 0 0 0.05 0 0 0.95 0.1 0 0 0 0 0 0 0.05 0.05Lead 0 0 0 0 0.05 0 0 0.05 0.95 0 0 0 0 0 0 0 0.05Be self 0 0 0 0 0 0 0 0.1 0.05 0.95 0 0 0 0 0 0 0.1Follow 0 0 0 0 0 0 0 0 0 0 0.95 0.05 0 0.1 0 0 0Mimic 0 0 0 0 0 0.05 0.1 0 0 0 0.05 0.95 0.05 0.05 0.05 0 0Be fair 0 0 0 0 0 0 0 0 0 0 0.05 0.05 0.95 0.05 0.05 0 0Follow code 0 0 0 0 0 0 0 0 0.05 0.05 0.1 0.1 0.1 0.95 0.01 0 0Be caring 0 0 0 0 0.1 0 0 0 0 0 0 0 0 0 0.95 0 0Organize 0 0 0 0 0 0.05 0 0 0 0 0.05 0.05 0 0 0 0.95 0Explore 0 0 0 0 0 0 0 0.05 0 0.05 0 0 0 0 0 0 0.95

Table 7List of actions (behaviors). The initials and the numbers in parenthesis are used for referencing figures later.

Eat/drink Stay at periphery Help others with work Ensure work distributed fairlyE/D (0) SP (11) HOW (22) EDF (33)Drink alcohol Self-disclose Order others what to do Wear something distinctiveDA (1) SD (12) OO (23) WSD (34)Relax Ask others about self Dance StealR (2) AO (13) D (24) S (35)Play practical joke Talk politics Ask other to dance Kiss upPPJ (3) TP (14) AOD (25) KU (36)Tease/make fun of Gossip/Talk About Others Ask for date Be cheapT/M (4) G/T (15) AD (26) BC (37)Try new dance steps Talk about work (job-related) Kiss MediateTND (5) TAW (16) K (27) M (38)Intro self to others Tell jokes Do job Give inISO (6) TJ (17) DJ (28) GI (39)Surf web Compliment others Extra effort job ProcrastinateSW (7) CO (18) EEJ (29) P (40)Explore environment Ignore others Find new way to do job Pretend to workEE (8) IO (19) FNJ (30) PW (41)Leave Insult others Improve skills Stay with comfortable othersL (9) InO (20) IS (31) SCO (42)Be silent Clean up Confront other about slackingBS (10) CU (21) COS (32)

chosen behavior on the y-axis (see Table 7 for the indices ofbehaviors).

We drilled down to see how individuals of different per-sonality types behaved differently in a given situation.Fig. 6 shows, as an example, the comparison between thesociable and the shy in the urgent project scenario. Ascan be seen from the figure, the sociable was more likelyto talk about work and help others, but less likely to stayat the periphery; the sociable and the shy were almostequally likely to put in extra effort (because this was anurgent project scenario); and so on.

As shown by these figures above, the model was adept atexhibiting appropriate behaviors of the different personal-ity types. An individual of a particular personality typeacting appropriately within a given situation was theresult of the interaction between the (relatively stable)

characteristics of the motivational and other subsystemsand the influence of the situation. For instance, the activa-tions of different drives were the results of stable internalmotivational parameters (such as the deficits of differentdrives), as well as stimuli received from situations. Further-more, which goal was pursued at any given moment waspartially a result of the competitive interaction amongdrives and which goal “won” that competition (stochasti-cally). Behaviors were then (stochastically) determinedbased on both the goal and the current situation.

4.3. Simulation 3

It is important to validate the CLARION personalitymodel against actual human data. In order to show thatCLARION could capture what simplified, specialized

Table 8The most plausible goal given a scenario and the generic personality.

Scenario Most plausible goal

Individual Assignment Be SelfWorking with 1 Other Fit inUrgent Project Fit inGroup Meeting MimicMeeting with Boss FollowTaking a Break with Coworkers MimicTaking a Break by Yourself Be SelfParty at Work Fit inSocial Engagement at Boss’s House Fit inDance MimicTrying to get a Date Be SelfOn a Date Fit inFamily Birthday Party MimicWedding Party at a Fancy Restaurant Fit inParty in a Restaurant that has a Bar Mimic

Table 9Accuracy of behaviors.

CLARION (%)

Top 1 89.8Top 3 78.2

Table 10Accuracy of behaviors with noise.

CLARION (%)

Top 1 86.1Top 3 78.1

personality models can capture, in this simulation, we didwhat Quek and Moskowitz (2007) did, that is, simulatingdata from Moskowitz, Suh, and Desaulniers (1994) andSuh, Moskowitz, Fournier, and Zuroff (2004).

Below we first examine data from Moskowitz et al.(1994) and Suh et al. (2004). Then simulations are pre-sented and compared to the simplified simulations by Quekand Moskowitz (2007). For additional validations, see thenext subsection, as well as prior papers such as Sun et al.(2011).

4.3.1. Human data

4.3.1.1. Human data from Moskowitz et al. (1994). Mosko-witz et al. (1994) examined the influence of social role oninterpersonal behavior. Moskowitz et al. hypothesized thatsocial roles would have a significant effect on an individ-ual’s behavior: Subjects were expected to behave more sub-missively when interacting with bosses vs. coworkers orsubordinates. Subjects were also expected to be more dom-inant with subordinates or coworkers than with bosses.

Subjects were asked to monitor their social interactionsfor 20 days. Subjects completed an event contingentrecording form for each significant social interaction

(defined as an interaction lasting at least 5 min). They wereasked to indicate on the form the gender, working relation-ship, and personal relationship of each person involved inthe social interaction. The form contained 46 behaviorsthat had been shown to be good indicators of dominance,submissiveness, agreeableness, and quarrelsomeness; sub-jects were asked to indicate which of the behaviors fromthe form they had engaged in during each significant socialinteraction.

We are interested in the findings related to dominant andsubmissive behaviors. In examining dominance, a significantmain effect was found for social role [F(2,102) = 5.33,p < .01]. Subjects reported more dominance toward superv-isees or co-workers than toward bosses, as confirmed withTurkey’s post hoc test (a < .05). In analyzing submissive-ness, there was also a significant main effect for social role[F(2,102) = 5.32, p < .01]. Subjects reported moresubmissiveness toward bosses than toward coworkers orsupervisees. The differences were significant by Turkey’spost hoc test (a < .05).

4.3.1.2. Human data from Suh et al. (2004). The experi-ment of Suh et al. (2004) was similar to Moskowitz et al.Subjects were asked to make event contingent recordingsabout non-work-related social interactions. On the eventcontingent reporting form, subjects were asked to specifyinformation about each interaction and the partner withwhom they had interacted: They were asked to indicatethe person’s gender and relationship as well as specify whatbehaviors took place during the interaction.

We are interested in the findings related to agreeable andquarrelsome behaviors. For agreeable behaviors, a signifi-cant Gender � Relationship interaction was found [F(2,5185) = 27.41, p < .001]. Agreeable behaviors with same-sex friends were significantly higher among women thanmen [t(5185) = 2.65, p < .01]. Agreeable behaviors withromantic partners were significantly higher among menthan women [t(5185) = 3.93, p < .001]. For quarrelsomebehaviors, a significant main effect for relationship wasfound [F(2, 5189) = 10.61, p < .001] as well as a significantGender x Relationship interaction [F(2,5185) = 8.24,p < .001]. Quarrelsome behaviors with romantic partnerswere significantly higher among women than men[t(5185) = 2.19, p < .05].

4.3.2. SimulationsSimulations of these two sets of human data above were

previously conducted by Quek and Moskowitz (2007).Their model was simple. A Backpropagation networkwas used. Only abstract scenarios were used. Instead ofspecific behaviors, Quek and Moskowitz simply set up theirmodel to choose between two outcomes (dominance vs.submissiveness for one simulation, agreeableness vs. quar-relsomeness for the other).

Our simulation was to add the necessary complexity andsophistication, on the basis of CLARION. Our simulationof the data was set up identically as our previous simulations,

Fig. 3. The most frequent behaviors of the sociable and the shy personality across the 15 scenarios. The Y-axis shows the indices of the most frequentbehaviors.

Fig. 4. The most frequent behaviors of the confident and the anxiouspersonality across the 15 scenarios.

Fig. 5. The most frequent behaviors of the responsible and the lazypersonality across the 15 scenarios.

except for necessary changes pertaining to the scenarios andthe behaviors. In particular, a subset of behaviors used pre-viously was involved here. They were classified as belongingto some of the four categories: dominance vs. submissive-ness, and agreeableness vs. quarrelsomeness, as can be seenin Table 11.10 Because in the real world, behaviors oftendid not fit neatly into the two classifications, each simulationalso included behaviors that did not fit into theclassifications.

4.3.2.1. Simulation of Moskowitz et al. (1994).

4.3.2.1.1. Simulation setup. Our simulation of Moskowitzet al. used some of the previous scenarios (slightly modifiedwith social roles added). The prior simulation of the same

10 The behaviors were categorized based on Moskowitz et al. (1994),which constructed an item pool for assessing behaviors in which itemswere divided into the experience-sampling scales for dominance, submis-siveness, agreeableness, and quarrelsomeness.

data by Quek and Moskowitz (2007) only specified 3 highlyabstract scenarios: the agent acting as a boss, the agent as acoworker, and the agent as a subordinate. To utilize ourearlier simulations, we used two earlier scenarios (urgent

project and work with 1 other). This leads to a total of 6variations; see Tables 12 and 13.

With these scenarios, after pre-training, each scenariowas tested for 100 steps, and the chosen behaviors wererecorded for each step. The process was repeated for 100runs representing 100 simulated “subjects”.4.3.2.1.2. Simulation results. As can be seen in Table 14, ourfindings were consistent with the human data of Mosko-witz et al. as well as the simulation of the data by Quekand Moskowitz.

As discussed before, Moskowitz et al. found that thatsubjects behaved more submissively when interacting witha boss vs. a coworker or a subordinate, and they behavedmore dominantly with subordinates or coworkers thanwith bosses. The same results were obtained from our sim-ulation, as shown by Table 14, the same as the (simplified)simulation by Quek and Moskowitz.

4.3.2.2. Simulation of Suh et al. (2004).

4.3.2.2.1. Simulation setup. The simulation of Suh et al.(2004) was the same as the previous simulation, except thatit involved non-work related scenarios. The four scenariosfor this simulation were as shown in Tables 15 and 16. Thesame features for coding the scenarios were used, exceptthat two features for indicating the gender of an individualwere added (the addition of features did not affect the over-all design, nor did any other parameters have to be altered).

Each scenario was tested for 100 steps, using the genericpersonality, and the chosen behaviors were recorded foreach step. The process was repeated for 100 runs represent-ing 100 simulated “subjects”.4.3.2.2.2. Simulation results. The findings from our simula-tion were consistent with the human data from Suh et al.,and with the simulation by Quek and Moskowitz. SeeTable 17.

According to Suh et al., when interacting with same-sexfriends, women exhibited significantly more agreeablebehaviors than men. When interacting with romantic

Stay at periphery

Talk about work (job-related)

Help others with work

Do job Find new way to do job

Ensure work distributed

fairly

Kiss up

Behaviors (included if >= 50 out of 10,000 for at least one of the personalities displayed)

Working together on urgent project: at work; conference room; in an office; conflict situation;

work to do; urgent; w/ 2 or more others; w/ subordinates; w/ disliked acquaintance

Sociable

job Extra effort Give in

Fig. 6. The sociable and the shy in the urgent project scenario. The Y-axis indicates the frequencies of the behaviors.

Table 11Two classifications of behaviors.

Classification 1 Classification 2

Dominance Submissiveness Neither Agreeableness Quarrelsomeness Neither

Order Others what to do(OO)

Be Silent (BS) Do Job (DJ) Tell Jokes (TJ) Insult Others (InO) Talk Politics (TP)

Confront Others aboutSlacking (COS)

Stay atPeriphery (SP)

Find New Ways to doJob (FNJ)

Compliment Others(CO)

Order Others what to do(OO)

Talk About Work(TAW)

Ensure Work DistributedFairly (EDF)

Kiss Up (KU) Kiss (K) Confront Others aboutSlacking (COS)

Insult Others (InO) Give In (GI) Give In (GI) Play Practical Joke (PPJ)Help Others with Work

(HOW)Leave (L) Ask Others about Self

(AS)Tease/Make Fun of (T/M)

Gossip/Talk aboutOthers (G/T)

Ignore Others (IO)

Table 12Scenarios for simulating Moskowitz et al. (1994).

Urgent Project as Boss:

urgent work; w/ subordinates; w/ 2 or more others; at work; work todo

Work with 1 Other as Boss:

w/ subordinates; w/ 1 other; at work; work to doUrgent Project as Coworker:

urgent work; w/ friends; w/ 2 or more others; at work; work to doWork with 1 Other as Coworker:

w/ friends; w/ 1 other; at work; work to doUrgent Project as Subordinate:

urgent work; w/ boss; w/ 2 or more others; at work; work to doWork with 1 Other as Subordinate:

w/ boss; w/ 1 other; at work; work to do

partners, men exhibited more agreeable behaviors and lessquarrelsome behaviors than women. Our simulation cap-tured all of the above, the same as Quek and Moskowitz.

4.3.3. DiscussionsIn some way, the two simulations of human data above

suggested some psychological validity of CLARION as apersonality model. Furthermore, they also suggested someplausible explanations for the data patterns observed inempirical studies; for example, different behaviors givendifferent roles were attributed to roles as inputs from situ-ations, rather than personality changes.

Because of the involvement of detailed representations,mechanisms, and processes (including drives, goals, andactions), the simulations provided deeper looks into thepsychological underpinning of the data concerningpersonality than the previous simplified simulations, in amechanistic, and process-based way, and suggested somepotentially useful explanations, which might be tested bylater empirical studies.

One shortcoming of the simulations above is the“tweaking” of parameters in the simulation setups (whichis the same as previous computational models such as

Read et al., 2010 and Quek & Moskowitz, 2007). Somemay regard parameter tweaking as being arbitrary. To rem-edy this problem, below we show that personality modelsmay be derived rigorously based on existing empirical(experimental and clinical) data, beyond what has beendone so far with other computational personality models.

4.4. Simulation 4

For further validating the CLARION personality model,a number of personality disorders (PDs) were simulated.Specifically, our objectives include: (1) demonstrating thatthe CLARION personality model may be rigorouslyderived from human data and behaves as predicted; person-ality disorders were chosen because of the availability ofusable human data related to personality disorders; (2)developing a motivation-centered approach to PDs, addingto our motivation-centered model of personality presentedso far. Below we will sketch some details of the simulationwithout being bogged down by clinical details (for furtherinformation, see Wilson & Sun, in preparation).

4.4.1. Empirical findings on personality disorders

A significant amount of empirical and clinical work hasbeen done on personality disorders. Such work led to someauthoritative documents categorizing personality disorders(as well as other mental disorders), such as DSM-IV (Diag-nostic and Statistical Manual of Mental Disorders, FourthEdition; American Psychiatric Association, 2000). Basedon DSM-IV, a number of diagnostic tools have been devel-oped, including SCID-II (Structured Clinical Interview ForDSM-IV-TR Axis II Personality Disorders) and PDQ-4(Personality Diagnostic Questionnaire for the DSM-IV).DSM-IV included ten different personality disorders. Wefocused on four of them: obsessive–compulsive personalitydisorder, schizotypal personality disorder, narcissistic per-sonality disorder, and avoidant personality disorder.

We discovered that there had been empirical work thatcorrelated measures of personality disorders and personal-ity measures (the Five-Factor Model known as NEO-PI-R), so that we could map PDs to personality measures(the Five-Factor Model). We also discovered that therehad been empirical work that correlated personality mea-sures and motivation measures, so that we could map per-sonality measures to motivations, and eventually to theCLARION drives. Such correlational studies included,for example, Samuel and Widiger (2008), Boyd (2010),and Costa and McCrae (1988). Using such data, we canmap PDs to the Five-Factor Model and then in turn tosome existing motivation measures, such as PRF (Person-ality Research Form) and RMP (Reiss Motivation Profile).PRF and RMP are very similar to the CLARION drivesdiscussed earlier, and therefore can be mapped to the drivesin a straightforward way. Thus, as a result of these map-pings above, PDs can be mapped to the CLARION drives.

The data found in correlational studies served as thebasis of our model of PDs. Our model thus provided a

Table 14Results of the simulation of Moskowitz et al. (1994). See the text for explanation.

CLARION simulation Quek and Moskowitz simulation Moskowitz et al.

Urgent Project Work with 1 Other

Dominance

Boss � Coworker t(1000) = 23.2 t(1000) = 35.42 t(3811) = 10.10 F(2,102) = 5.33Boss � Subordinate p < .001 p < .001 p < .001 p < .01

t(1000) = 92.32 t(1000) = 246.39 t(1395) = 6.28 F(2,102) = 5.33p < .001 p < .001 p < .001 p < .01

Submissiveness

Subordinate � Boss t(1000) = 97.58 t(1000) = 93.38 t(1395) = 8.80 F(2,102) = 5.32p < .001 p < .001 p < .001 p < .01

Coworker � Boss t(1000) = 29.06 t(1000) = 25.48 t(3811) = 7.58 F(2,102) = 5.32p < .001 p < .001 p < .001 p < .01

Table 15Scenarios for simulating Suh et al. (2004).

Male with Same Sex Friend:

w/ same-sex friends; w/ 1 other; maleFemale with Same Sex Friend:

w/ same-sex friends; w/ 1 other; femaleMale with Romantic Partner:

w/ 1 other; w/ romantic partner; maleFemale with Romantic Partner:w/ 1 other; w/ romantic partner; female

motivation-centered account for personality disorders (andat the same time a computational validation for thesecorrelational studies), from a cognitive architecture per-spective. The model was also constructed based on theinformation concerning symptomatic behaviors, typicalscenarios that elicit symptomatic behaviors, and so on,from SCID-II, PDQ-4, and DSM-IV. Then, the completemodel was tested and validated using PDQ-4 and SCID-II (for diagnosing PDs). Our model was derived strictlyon the basis of these documents without arbitrary parame-ter tweaking.

4.4.2. Simulating personality disorders

4.4.2.1. Simulation setup. For the sake of simulation, fourgroups of simulated “subjects” with four different typesof PDs respectively were formed, including obsessive–com-pulsive, schizotypal, narcissistic, and avoidant personalitydisorders. A non-PD group was also included in simulationas the baseline. 10 simulated “subjects” were included ineach group. Model parameters for each group were deter-mined based on well-established clinical and other docu-ments, thus avoiding arbitrary parameter tweaking.

It was hypothesized that personality disorders might be(mostly) explained by drive-related processes (because theyare more stable and less transient; Principle 5), while somemore transient types of mental disorders might be morereadily explained by other processes (Sun et al., 2011).

The model for this simulation was based on thesimulations earlier. The ACS involved a pre-trainedBackpropagation network at the bottom level, with 67

input nodes (for 44 goals and 23 scenarios), 108 outputnodes (for 108 behaviors), and 85 hidden nodes.

Behaviors symptomatic for each PD were determinedbased on the information from DSM-IV as well as theother related documents. This process led to a total of108 behaviors. For example, see Table 18 for the symp-tomatic behaviors of OCPD and their documentarysources.

For constructing scenarios that would elicit thesebehaviors, the key evaluation criteria in the SCID-II man-ual (Section 4.2-B) were used, and one or more situationswere selected for each criterion. Scenarios were selectedwhere symptomatic behaviors were likely to occur as sug-gested by the SCID-II manual. This led to a total of 23scenarios as shown in Table 19. Again using OCPD asan example, Table 20 shows possible occurrences of thesymptomatic behaviors (as identified in Table 18) in thesescenarios.

For a motivation-centered account of PDs, we need amapping from PDs to the drives within the MS of CLAR-ION. Three steps were devised to calculate this mapping:

(1) Mapping PDs to the Five-Factor Model of personal-ity (NEO-PI-R). We based the mapping on thehuman data from Samuel and Widiger (2008,Table 3).

(2) Mapping the Five-Factor Model (NEO-PI-R) to theCLARION drives: We devised three separate meth-ods in this regard:

sonality Research Form) then to the CLAR-ION drives, by using the correlation datafrom Boyd (2010, Table 9) and Costa and McC-rae (1988).

(2b) Mapping the Five-Factor Model to RMP (ReissMotivation Profile) then to the CLARIONdrives, by using the correlation data from Boyd(2010, Table 5).

(2c) Mapping the Five-Factor Model to PRF (Per-sonality Research Form), then to RMP (ReissMotivation Profile), and then to the CLARION

The following considerations were also used in mapping lx,z to valuesdrive deficits: :40 < ‘x;z < :60 were excluded (which corresponded to

correlations). Reiss Motivation Profile (i.e., methods 2b and 2c) wasferred over Personality Research Form (i.e., method 2a) (because of theter mapping between Reiss Motivation Profile and the CLARIONves; Sun, 2009). Contradictions were excluded (e.g., when methods 2bd 2c suggested opposite deficits: low vs. high). Drives recommendedly by method 2a were excluded (because of the less reliable mappingween Personality Research Form and CLARION drives).

Asbyinpapdrsugivquwafrotiomago

oflowprebetdrianonbet

drives, by using correlation data from Boyd(2010, Tables 8 and 9) and Costa and McCrae(1988). Successive correlation measures aremultiplied along the way.

3) Mapping the afore-obtained correlations to the drivedeficits in CLARION: The mapping from PDs todrive deficits was calculated, by combining the corre-lation from the PDs to the Five-Factor Model andthe correlation from the Five-Factor Model to thedrive deficits (based on one of the three methodsabove):

max�min

where x is any of the four PDs, y is any facet of theFive-Factor Model, z is any drive, m and n are corre-lations obtained from the two steps above, max = 1,and min = �1.11

The final value of the deficit for a drive was set byaveraging the values lx,z obtained from each of thethree methods above. Drive stimulus values were setat 1. All gain parameters were set to 1.

Finally, goal setting was accomplished based on drives.discussed earlier, each goal should be jointly determinedthe current drive activations and the current sensoryuts (denoting scenarios), so that it would be context-

propriate. Specifically, the goal was set based on “high”ives, where a drive was “high” if its deficit P.6. For eachch drive, a goal would be identified that was appropriate,en the scenario and the drive, considering the PD inestion. The relevance parameter of the drive to the goals then set to 1 (see the equation earlier). This mappingm drives to goals was developed using the drive defini-n of Sun (2009) and the information from the SCID-IInual regarding PDs. This process led to a total of 44

als (see Table 21).All other parameters of the model were the same as in

e previous simulations.

.2.2. Simulation results. The resulting models were sub-ted to the two standard PD diagnostic tests: SCID-IId PDQ-4. Figs. 7–10 show the results of the tests.The simulated “subjects” in the OCPD group (see theulation setup earlier) obtained a score of 7.00 from bothPD tests of the SCID-II and the PDQ-4, well above the

Table 17Results of the simulation of Suh et al. (2004). See the text for explanation.

CLARION simulation Quek and Moskowitz simulation Suh et al.Men vs. Women Men vs. Women Men vs. Women

Interactions with same-sex friends

Agreeableness t(1000) = 58.35 t(2041) = 2.04 t(5185) = 2.65p < .001 p < .001 p < .01

Interactions with romantic partners

Agreeableness t(1000) = 28.46 t(2060) = 6.83 t(5185) = 3.93p < .001 p < .001 p < .001

Quarrelsomeness t(1000) = 31.9 t(2060) = 6.64 t(5185) = 2.19p < .001 p < .001 p < .05

Table 18Symptomatic behaviors of OCPD.

DSM-IV Symptom # SCID-II Question # PDQ-4 Question # Behaviors

OCPD-1 16 3 OrganizeOCPD-1 16 3 PlanOCPD-1 16 3 ScheduleOCPD-2 17 16 Insist things be exactly rightOCPD-6 21 66 Refuse to delegate tasksOCPD-4 19 41 ChastiseOCPD-3 18 29 Get ahead on workOCPD-3 18 29 Get caught up on workOCPD-5 20 54 HoardOCPD-7 22 81 SaveOCPD-8 23 89 Insist on rightness of selfOCPD-8 24 89 Be rigid about opinions

Table 19Scenarios for testing various PDs.

Scenarios (based on SCID-II) OCPD Schizotypal Narcissistic Avoidant

Learning how to do something new U

Project at work U U U

Social situation (e.g., party) U U U

With spouse/lover U U U

With friends/confidants U U U

With coworkers U U U U

With family U U

Meeting new people U U

Being asked to talk about self U U

Being asked to attend an activity w/ others U U U

In situation requiring empathy w/ others U

In situation where others are being recognized U

In situation requiring hurdles be overcome in order to advance U

Situation where money may be spent U

Household cleaning task U

Going on a trip U

Situation where unlikely outcome desired (e.g., winning lottery) U

In public near strangers U

Dressing for a hot day U

Front desk assignment U

Working in a coffee shop U

Looking for perfect romance U

Waiting in line for something U

Table 20Possible occurrences of the symptomatic behaviors of OCPD in some scenarios.

DSM IVSymptom #

SCID-IIQuestion #

PDQ-4Question #

Behaviors Project atwork

Withfriends. . .

Withcoworkers

Withfamily

. . .money maybe spent

. . .cleaningtask

. . .ontrip

1 16 3 Organize U U U

1 16 3 Plan U U

1 16 3 Schedule U

2 17 16 Insist. . .exactlyright

6 21 66 Refuse todelegate. . .

4 19 41 Chastise U U U

3 18 29 Get ahead onwork

3 18 29 Get caught upon work

5 20 54 Hoard U U U U

7 22 81 Save U U U U

8 23 89 Insist onrightness. . .

8 24 89 Be rigid. . . U U U U U

12 It should be pointed out that the simulations so far relied mostly onneural networks at the bottom level of CLARION. Other existingcognitive architectures that rely heavily on neural networks are also likelyto be applicable in this regard.

threshold of 4 for both tests. Therefore these simulated“subjects” were indeed OCPD.

The simulated “subjects” in the avoidant groupobtained a score of 5.00 from both avoidant tests of theSCID-II and the PDQ-4, while the threshold in both caseswas 4. Therefore the simulated “subjects” were indeedavoidant.

The simulated “subjects” in the schizotypal groupobtained a score of 5.66 from the schizotypal test ofSCID-II and a score of 6.00 from the schizotypal test ofPDQ-4, while the threshold in both cases was 5. Thereforethe simulated “subjects” were indeed schizotypal.

Finally, the simulated “subjects” in the narcissisticgroup obtained a score of 6.50 from the narcissistic testof SCID-II and a score of 7.00 from the narcissistic testof PDQ-4, while the threshold in both cases was 5. There-fore the simulated “subjects” were determined to benarcissistic.

The results validated our CLARION-based model ofpersonality disorders. There are of course many furthertheoretical or technical questions one may ask concerningthis approach, which, however, are outside of the scopeof the present paper. Due to length and other consider-ations, they cannot be addressed here (for further discus-sions and details, see Wilson & Sun, in preparation).

5. General discussions

5.1. Assessment

The present CLARION model of personality is adetailed and precise computational model, involving themotivational, metacognitive, procedural, and otherprocesses. It captures processes underlying personality, sothat as the external situation and the internal state changeover time, it generates behaviors that vary. At the same

time, the structure and parameters of the subsystems cancapture relatively invariant, stable individual differencesin behavioral inclinations and tendencies, namely personal-ity traits. That is, the present model can provide an accountof broad, stable traits, as well as an account of the under-lying processes that lead to behavioral variability across sit-uations and over time. In particular, it relies on dualrepresentations in various subsystems (i.e., drives andgoals, as well as dual metacognitive, dual procedural, anddual declarative representations).12

In the present model, person–situation interaction(Caprara & Cervone, 2000) is the interaction between the(relatively invariant) characteristics of the subsystems andthe influence of situations (which are more transient).Within our model, we can easily vary either personalityor situation (or both). Thus both factors may be tested,separately or in conjunction. Note that individual differ-ences (in terms of the parameters in the model) might be,in part, attributed to innate (“hardwired”) differences(e.g., due to genetic factors), and also, in part, to differentindividual experiences during ontogenesis (including differ-ent experiences of sociocultural influences; see Sun (2003)regarding learning).

We have shown that CLARION, as is, can serve as ageneric model of personality. In fact, it is broader and morecomprehensive than other computational models of per-sonality. It can account for whatever phenomena those(more specialized) computational personality models canaccount for. Beyond that, it can also account for phenom-ena that those computational models have not been able toaccount for thus far. In particular, the personality disorder

Table 21Goal setting based on situations and drives.

(continued on next page)

Table 21 (continued)

tests that our model undergoes serve as a form of “person-ality Turing test”. Notably, the present model may captureempirical data without arbitrary tweaking of parameters(e.g., simulation 4). The present model integrates a numberof strands of psychology and cognitive science, includingsome of personality psychology and cognitive architectureresearch.

While still far from fully establishing the validity of theCLARION personality model, the simulation results abovedo suggest, in a preliminary way, some psychological valid-ity of the model. Development and validation of computa-tional models are difficult. The present work is only aninitial attempt in this regard; much more work needs tobe done in the future. The model is not yet comprehensivewith regard to all issues covered by personality psychology,but might provide the basis for a future comprehensivepersonality model (including motivation, emotion, meta-cognition, coping, self-theory, instinct, intuition, and soon; Sun & Mathews, 2012), due to the foundation of abroadly scoped cognitive architecture.

5.2. Objections

One possible line of objection may claim that personal-ity has been built into CLARION and thus it is no surprisethat it accounts for personality. In this regard, it should beemphasized that the motivational subsystem in CLARIONwas based on a line of research that is quite different fromthe FFM and other contemporary personality theories.Our work shows, among other things, the connectionbetween these different lines of work, which may be usefulin its own right. Even given drives (such as “power anddomination”), one would still need to show how they leadto relevant behaviors in relevant situations, with goalsetting, action selection, and possibly reasoning as well.Our model demonstrates these steps with processes thathave been tested previously through simulating a widevariety of psychological data in different domains.

Another possible line of objection may claim that themotivational subsystem is not adequately justified (e.g.,“what about cultural variability?”, “what about subdividing

Fig. 7. Test results of the simulated OCPD group.

these drives?”, and so on). As noted before, our primarydrives are essentially the same as Murray’s needs (1938),with only a few differences (e.g., conservation in our set ofdrives covers both the need for conservance and the needfor order proposed by Murray, and so on). Similarly, com-paring this set of drives with Reiss’ motives (2010), one cansee that they are highly similar (but with some differences).In addition, Schwartz’s (1994) 10 universal values, althoughaddressing a different issue, bear some resemblance to thedrives identified here. Each of his values can be derived fromsome primary drive or some combination of them. So, theprior work by these researchers in justifying their frame-works may be applied, to a significant extent, to our drivesas well (Maslow, 1943; McDougall, 1936; Murray, 1938;Reiss, 2010). Moreover, this work is at the psychologicallevel, and thus does not make any claim regarding imple-mentations of these drives (e.g., neurophysiology). In addi-tion, these drives appear to be (more or less) innate anduniversal across cultures (while their expression may varyacross cultures). See more justifications in Sun (2009).

Yet another possible issue is how one generatescombinations of personality types. For example, from therepresentations of two existing types “shy” and “responsi-ble”, one would want to be able to represent “shy andresponsible” as well. In this regard, note that personalityis, in part, the result of interactions among different drives(among other things), and that there may not necessarily bea direct relationship between the characteristics of a singledrive (or a single group of drives) and the correspondingtrait (Smillie et al., 2006). Furthermore, the drive parame-ter settings for those types used for simulation earlier (suchas “shy” or “responsible”) should be viewed as prototypicalsettings, and they may change when generating combina-tion traits (such as “shy and responsible”). Thus the modelis fully capable of dealing with combinations. In addition,new personality types may emerge through differentsettings of parameters, and possibilities are abundant.

Finally, another claim is that our work has been influ-enced by previous, more specialized computational mod-els of personality, such as Shoda and Mischel (1998),

Fig. 8. Test results of the simulated avoidant group.

13 As indicated before, a cognitive architecture should have essentialcomponents of the mind (such as various memory modules, goals, andinference mechanisms) built-in. Otherwise it amounts to a software tool.Such a software tool was used by Read et al. (2010).

Quek and Moskowitz (2007), or Read et al. (2010). This iscertainly true, especially in relation to simulation setups.However, as we noted earlier, personality did emerge fromCLARION without requiring significant additions ormodifications of CLARION, which speaks to the relevanceof cognitive architectures for personality research.

5.3. Comparisons

A comparison of the CLARION personality model withexisting (specialized) computational models of personalityis in order. Compared with these other models, which weredesigned to address only personality, the CLARION per-sonality model has the following general advantages: (1)it is contextualized within a relatively generic and compre-hensive cognitive architecture, (2) it is psychologically andcomputationally detailed (as required by the cognitivearchitecture), (3) it is grounded in a theory of essentialhuman motivation (as contained within the cognitive archi-tecture), (4) it connects with existing psychological dataand theories. We examine a few of these previous modelsin more detail below.

First, CLARION differs from Read et al. (2010). Theirmodel was goal-based. Some dimensions of the Five-

Factor Model were mapped into goal structures (althoughno detailed facets of the FFM were used). Read et al. useda complex, biologically inspired neural network model.13

There are some similarities between their model and thepresent model (e.g., reliance on motivation). But thereare some significant differences, because CLARION wasbased on different research traditions (and CLARIONhas a longer history). For instance, (1) the present modelis based on the foundation of a generic but detailed cogni-tive architecture (which has undergone development forover two decades using a variety of psychological data),and thus based on a more comprehensive and bettergrounded view of the architecture of the mind, which hasled to unified explanations of a wide range of psychologicalphenomena (together with personality phenomena). (2)The present model captures personality types based onboth drives and goals, in which drives serve as the basisfor setting goals. (3) The present model is based on a moti-vational subsystem that synthesized well-developed past

Fig. 9. Test results of the simulated schizotypal group.

theories on motivational processes (e.g., Murray, 1938;Reiss, 2010; Sun, 2009). (4) The present work suggests thattranslation from personality types to drive activations canbe done in an empirically grounded, systematic, non-arbi-trary way, minimizing parameter tweaking. (5) In addition,CLARION dealt with personality disorders along with per-sonality structures.

Shoda and Mischel (1998) developed a constraint satis-faction model of personality based on the cognitive sociallearning theory of personality. In their model, which wasa connectionist network, a set of input units (representingsituational features) was connected to a recurrently con-nected set of mediating cognitive affective units, which werethen connected to behavior units. Although their modelwas aimed to represent human personality, they did notattempt to capture established structures of personality(e.g., the Five-Factor Model). In their model, different per-sonalities are captured by randomly connected patterns ofmediating units, and they did not specify the structure ofthe underlying motivational subsystem. In contrast, thepresent work demonstrates that the structures of humanpersonality and personality disorders may be the result ofthe complex dynamics of the underlying motivational andcognitive processes.

Quek and Moskowitz (2007) used a simple Backpropa-gation neural network to simulate empirical data fromevent contingent recording. The simulation involved three

input nodes corresponding to three possible roles in theworkplace (supervisor, co-worker, or supervisee), a hiddenlayer, and two output nodes (for two types of behavior,respectively). The model learned the relationship betweenwork role and dominant vs. submissive behavior. Inanother simulation, it learned the relationship between gen-der and behavior in different kinds of relationships.Although the network could capture the empirical relation-ships, it did not include realistic motivational and cognitiveprocesses. It pre-coded only the relevant variables andignored contextual factors. A more realistic simulationwould need to take into account various contextual factorsand detailed motivational and cognitive processes, as hasbeen attempted in the present work.

Schaub (2001) and Doerner (2003) focused on modelingpersonality-specific behavior in specific situations. Theirbasic constructs included internal deficits, displeasure sig-nals, negative reinforcement (from displeasure signals),urges, goals, action learning through random exploration,and so on. The variation of parameter settings in theirmodel generated different “personality” patterns, whichdiffered in the way of coping with specific situations. Tosome extent, they showed that the behavior patterns pro-duced by the model were consistent with empirical data(although at a very abstract level). However, the downsideincluded the lack of detailed psychological justifications ofthe model.

Fig. 10. Test results of the simulated narcissistic group.

5.4. Future work

There are many possibilities for further extending thepresent work. For instance, future work on CLARIONmay include:

� Accounting for and simulating a wider variety of empir-ical data concerning human personality.� Conducting human experiments for exploring parame-

ters of human personality on the basis of the CLARIONpersonality model (its parameters and mechanisms).� Further exploring the potential of accounting for vari-

ous types of mental disorders (including personality dis-orders and beyond).� Capturing and incorporating a wider variety of related

psychological phenomena and processes.� Applications to cognitive social simulation, digital sto-

rytelling, synthetic characters, and many more areas(Sun, 2006, 2012).

Much of such work is already underway (e.g., Wilson &Sun, in preparation).

Acknowledgment

This work has been supported in part by the ONRGrants N00014-08-1-0068 and N00014-13-1-0342.

References

Aarts, H., & Hassin, R. (2005). Automatic goal inference and contagion.In J. Forgas, K. Williams, & S. Laham (Eds.), Social motivation:

Conscious and unconscious processes. New York: Cambridge UniversityPress.

American Psychiatric Association (2000). Diagnostic and statistical manual

of mental disorders (4th ed.). Arlington, VA.Anderson, J. A., & Lebiere, C. (1998). The atomic components of thought.

Mahwah, NJ: Lawrence Erlbaum Associates.Boyd, S. E. (2010). A comparison of the Reiss profile with the NEO PI-R

assessment of personality. Master’s thesis, University of Kentucky,Kentucky, USA.

Caprara, G. V., & Cervone, D. (2000). Personality: Determinants,

dynamics, and potentials. New York: Cambridge University Press.Castelfranchi, C. (2001). The theory of social functions: Challenges for

computational social science and multi-agent learning. Cognitive

Systems Research, 2(1), 5–38 (special issue on the multi-disciplinarystudies of multi-agent learning (Ed. Ron Sun)).

Cervone, D. (2004). The architecture of personality. Psychological Review,

111(1), 183–204.Cleeremans, A., Destrebecqz, A., & Boyer, M. (1998). Implicit learning:

News from the front. Trends in Cognitive Sciences, 2(10), 406–416.Costa, P. T., & McCrae, R. R. (1988). From catalog to classification:

Murray’s needs and the five-factor model. Journal of Personality and

Social Psychology, 55(2), 258–265.Curran, T., & Keele, S. (1993). Attention and structure in sequence

learning. Journal of Experimental Psychology: Learning, Memory, and

Cognition, 19, 189–202.Deci, E. (1980). Intrinsic motivation and personality. In E. Staub (Ed.),

Personality: Basic issues and current research (pp. 35–80). EnglewoodCliffs, NJ: Prentice Hall.

Doerner, D. (2003). The mathematics of emotions. In D. D. Frank Detje,& H. Schaub (Eds.), Proceedings of the fifth international conference on

cognitive modeling, Bamberg, Germany (pp. 75–79).Elliot, A., & Thrash, T. (2002). Approach–avoidance motivation in

personality: Approach and avoidance temperaments and goals. Jour-

nal of Personality and Social Psychology, 82(5), 804–818.Epstein, A. (1982). Instinct and motivation as explanations for complex

behavior. In D. W. Pfaff (Ed.), Physiological mechanisms of motivation.New York: Springer-Verlag.

Flavell, J. (1976). Metacognitive aspects of problem solving In: B. Resnick.The Nature of Intelligence. Hillsdale, NJ: Erlbaum.

Hall, C. S., & Gardner, L. (1985). Introduction to the theories of

personality. Toronto: John Wiley & Sons.Helie, S., & Sun, R. (2010). Incubation, insight, and creative problem

solving: A unified theory and a connectionist model. Psychological

Review, 117, 994–1024.Hull, C. (1943). Principles of Behavior: An Introduction to Behavior

Theory. D. Appleton-Century Company.Hull, C. (1951). Essentials of Behavior. New Haven, CT: Yale University

Press.John, O. P., & Srivastava, S. (1999). The Big Five trait taxonomy: History,

measurement, and theoretical perspectives. In L. A. Pervin & O. P.John (Eds.), Handbook of personality: Theory and research (2nd ed.,pp. 102–138). New York: Guilford Press.

Karmiloff-Smith, A. (1986). From meta-processes to conscious access:Evidence from children’s metalinguistic and repair data. Cognition, 23,95–147.

Maddi, S. R. (1996). Personality theories: A comparative analysis (6th ed.).Toronto: Brooks/Cole Publishing Co.

Maslow, A. (1943). A theory of human motivation. Psychological Review,

50, 370–396.Mayer, J. D. (2005). Tale of two visions: Can a new view of personality

help integrate psychology?. American Psychologist 60(4), 294–307.McAdams, D. P., & Pals, J. L. (2006). A new Big Five: Fundamental

principles for an integrative science of personality. American Psychol-

ogist, 61(3), 204–217.McDougall, W. (1936). An introduction to social psychology. London:

Methuen & Co..Montague, P. R. (1999). Review of reinforcement learning: An introduc-

tion. Trends in Cognitive Science, 3(9), 360–361.Moskowitz, D. S., Suh, E. J., & Desaulniers, J. (1994). Situational

influences on gender differences in agency and communion. Journal of

Personality and Social Psychology, 66, 753–761.Murray, H. (1938). Explorations in personality. New York: Oxford

University Press.Pollock, J. (2008). How to build a person: A prolegomenon. MIT Press.Over, H., & Carpenter, M. (2009). Eighteen-month-old infants show

increased helping following priming with affiliation. Psychological

Science, 20, 1189–1193.Quek, M., & Moskowitz, D. S. (2007). Testing neural network models of

personality. Journal of Research in Personality, 41, 700–706.Read, S. J., Monroe, B. M., Brownstein, A. L., Yang, Y., Chopra, G., &

Miller, L. C. (2010). Virtual personalities II: A neural network modelof the structure and dynamics of human personality. Psychological

Review, 117, 61–92.Reber, A. (1989). Implicit learning and tacit knowledge. Journal of

Experimental Psychology: General., 118(3), 219–235.Reder, L. (Ed.). (1996). Implicit memory and metacognition. Mahwah, NJ:

Erlbaum.Reiss, S. (2010). Skinny on Maslow’s hierarchy: Is Maslow’s hierarchy

valid? Psychology Today (July 24).Rumelhart, D., McClelland, J., & the PDP Research Group (1986).

Parallel distributed processing: Explorations in the microstructures of

cognition. Cambridge, MA: MIT Press.Ryckmann, R. M. (1993). Theories of personality (5th ed.). California:

Brooks/Cole Publishing Co.Samuel, D. B., & Widiger, T. A. (2008). A meta-analytic review of the

relationships between the five-factor model and DSM-IV-TR person-

ality disorders: A facet level analysis. Clinical Psychology Review,

28(8), 1326–1342.Schaub, H. (2001). The anatomy of human personality: A computational

implementation. In: Proceedings of the fourth international conference

on cognitive modeling (ICCM-2001). Fairfax, Virginia.Schwartz, S. (1994). Are there universal aspects of human values? Journal

of Social Issues, 50, 19–45.Schopenhauer, A. (1819). Die Welt als Wille und Vorstellung (The world as

will and representation. E.F.J. Payne, Trans. New York: DoverPublications, 1969.).

Seger, C. (1994). Implicit learning. Psychological Bulletin, 115(2), 163–196.Shoda, Y., & Mischel, W. (1998). Personality as a stable cognitive–

affective activation network: Characteristic patterns of behaviorvariation emerge from a stable personality structure. In S. J. Read &L. C. Miller (Eds.), Connectionist models of social reasoning and social

behavior (pp. 175–208). Mahwah, NJ: Lawrence Erlbaum AssociatesInc.

Simon, H. (1967). Motivational and emotional controls of cognition.Psychological Review, 74, 29–39.

Smillie, L. D., Pickering, A. D., & Jackson, C. J. (2006). The newreinforcement sensitivity theory: Implications for personality measure-ment. Personality and Social Psychology Review, 10, 320–335.

Staub, E. (1980). Defining personality. In E. Staub (Ed.), Personality:

Basic issues and current research (pp. 2–33). Englewood Cliffs, NJ:Prentice Hall.

Suh, E. J., Moskowitz, D. S., Fournier, M., & Zuroff, D. C. (2004).Gender and relationships: Influences on agentic and communalbehaviors. Personal Relationships, 11, 41–59.

Sun, R. (1994). Integrating Rules and Connectionism for Robust Common-

sense Reasoning. New York: John Wiley and Sons.Sun, R. (2002). Duality of the mind: A bottom-up approach toward

cognition. Mahwah, NJ: Lawrence Erlbaum Associates.Sun, R. (2003). A tutorial on CLARION 5.0. Technical report. Cognitive

Science Department, Rensselaer Polytechnic Institute. <http://www.cogsci.rpi.edu/rsun/sun.tutorial.pdf>.

Sun, R. (2004). Desiderata for cognitive architectures. Philosophical

Psychology, 17, 341–373.Sun, R. (2006). Cognition and multi-agent interaction: From cognitive

modeling to social simulation. New York: Cambridge University Press.Sun, R. (2009). Motivational representations within a computational

cognitive architecture. Cognitive Computation, 1(1), 91–103.Sun, R. (Ed.). (2012). Grounding social sciences in cognitive sciences.

Cambridge, MA: MIT Press.Sun, R., & Mathews, R. (2012). Implicit cognition, emotion, and meta-

cognitive control. Mind and Society, 11(1), 107–119.Sun, R., Merrill, E., & Peterson, T. (2001). From implicit skills to explicit

knowledge: A bottom-up model of skill learning. Cognitive Science, 25,203–244.

Sun, R., Slusarz, P., & Terry, C. (2005). The interaction of the explicit andthe implicit in skill learning: A dual-process approach. Psychological

Review, 112, 159–192.Sun, R., & Wilson, N. (2011). Motivational processes within the

perception–action cycle. In V. Cutsuridis, A. Hussain, & J. G. Taylor(Eds.), Perception–action cycle: Models, architectures, and hardware

(pp. 449–472). Berlin: Springer.Sun, R., Wilson, N., & Mathews, R. (2011). Accounting for certain mental

disorders within a comprehensive cognitive architecture. Cognitive

Computation, 3(2), 341–359.Taylor, C. (1985). The concept of a person. Philosophical papers (Vol. 1).

Cambridge: Cambridge University Press.Toates, F. (1986). Motivational systems. Cambridge, UK: Cambridge

University Press.Tolman, E. C. (1932). Purposive behavior in animals and men. New York:

Century.Tyrell, T. (1993). Computational mechanisms for action selection. Ph.D.

thesis. University of Edinburgh, Edinburgh, UK.Watkins, C. (1989). Learning with delayed rewards. Ph.D thesis. Cam-

bridge University, Cambridge, UK.

Weiner, B. (1992). Human motivation: Metaphors, theories, and research.Newbury Park, CA: Sage.

Wiggins, J. (Ed.). (1996). The five-factor model of personality. The GuilfordPress.

Wilson, N., & Sun, R. (in preparation). A model of personality disorders.Winter, D. G., John, O. P., Stewart, A. J., Klohnen, E. C., & Duncan, L.

E. (1998). Traits and motives: Toward an integration of two traditionsin personality research. Psychological Review, 105(2), 230–250.

Wood, W., & Quinn, J. (2005). Habits and the structure of motivation ineveryday life. In J. Forgas, K. Williams, & S. Laham (Eds.), Social

motivation: Conscious and unconscious processes. New York: Cam-bridge University Press.

Wright, I. P., & Sloman, A. (1997). MINDER1: An implementation of a

proto-emotional agent architecture. Technical report CSRP-97-1. Uni-versity of Birmingham, School of Computer Science. <ftp://ftp.cs.bha-m.ac.uk/pub/tech-reports/1997/CSRP-97-01.ps.gz>.

A model of personality should be a cognitive architecture itself

Documents

A social cognitive theory of personality

Identification Problems in Personality Psychologyftp.iza.org/dp5605.pdf · · 2011-04-05Identification Problems in Personality Psychology* ... Personality traits and cognitive abilities,

Behavioral Approach to Personality Social Cognitive Theory Says personality is shaped by 3 forces: environmental conditions, cognitive- personal factors,

Personality Social Cognitive approach. Social Cognitive- Bandura understanding personality involves considering the situation and thoughts before, during,

Tom Farsides ATP PAID 3: Socio-Cognitive Aspects of Personality Socio-Cognitive Aspects of Personality

PERSONALITY TRAIT AND COGNITIVE ABILITY CORRELATES …

BS 7 OTHER PSYCHIATRIC DISORDERS Cognitive disorders Personality disorders Dissociative disorders Obesity eating disorders Cognitive disorders Personality

Personality and cognitive differences between online and

anxiety, cognitive, somatoform, personality

Impact of narcissistic personality disorder on cognitive

Personality, Motivation, and Cognitive Performance

Personality What is your personality?. What are the ideas about personality? Psychoanalytic Humanistic Trait Social cognitive The self

SOCIAL COGNITIVE THEORY OF PERSONALITY - …eushe2/BanduraPubs/Bandura1999HP.pdf · · 2012-10-04SOCIAL COGNITIVE THEORY OF PERSONALITY Albert Bandura ... the very substance of

Personality and Sport. What Is Personality? Measuring Personality Cognitive Strategies and Athletic Success What Makes Up Personality? Approaches to Understanding

gender and personality differences in cognitive tasks

Cognitive Therapy of Personality Disorders - …Aaron_T__Beck... · Cognitive Therapy of Personality Disorders SECOND EDITION Aaron T. Beck Arthur Freeman Denise D. Davis and Associates

Cognitive abilities, personality and interests

Behaviorism Personality Theory, also known as Cognitive-Social Personality Theory. Albert Bandura Walter Mischel

Personality Assessment. What is Personality? Aiken (2003) defines personality as “a composite of cognitive abilities, interests, attitudes, temperament,

Social Cognitive Theory of Personality