Integrating Neural Networks into Decision-Making

Integrating Neural Networks into Decision-Making and Motivational Theory: Rethinking VIE Theory* ROBERT G. LORD

University of Akron PAUL J. HANGES ELLEN G. GODFREY

University of Maryland

Abstract This manuscript uses a reformulation of Vroom’s (1964) VIE theory to illustrate the potential value of neuropsychologically based models of cognitive processes. Vroom’s theory posits that people’s decisions are determined by their affective reactions to certain outcomes (valences), beliefs about the relationship between actions and outcomes (expectancies), and perceptions of the association between primary and secondary outcomes (instrumentalities). One of the major criticisms of this type of theory is that the computations it requires are unrealistically timeconsuming and often exceed working memory capacity. In this paper, we maintain that if an individual has extensive experience with a problem situation, he or she can process decisions about that situation using neural networks that operate implicitly so that cognitive resources are not exhausted by simple computations; instead, the computations are performed implicitly by neural networks. By thinking about VIE from a neural network standpoint, at least one of its problems is eliminated, and several new insights into decision-making are provided. We use simulation methodology to show that such a model is both viable and can reflect the effects of current goals on choice processes.

The transformation of social psychology by the information processing metaphor in the 1970s and 1980s affected applied psychology in a number of ways, influencing our understanding of critical applied processes such as performance appraisal (Ilgen, Barnes-Farrell, & McKellin, 1993) and leadership perceptions (Lord, *This article was accepted by the previous editor, Dr. Vic Catano/Cet article a été accepté par le rédacteur en chef précédent, Dr Vic Catano. Canadian Psychology/Psychologie canadienne, 44:1 2003

Foti, & De Vader, 1984). Social psychology is currently being transformed by another metaphor in which the analogy to brain structure and processes serves to constrain both cognitive and social theory. This social-cognitive-neuroscience perspective uses cognitive processes as a key linkage between the social and neuropsychological levels of theory development (Ochsner & Lieberman, 2001). Applied psychologists, however, have rarely ventured into the area of neuroscience, in part because the methodologies of neuroscience are seen as being too intrusive for applied use and because the neuroscience knowledge base is generally unfamiliar to practitioners. In the current article, we attempt to illustrate the potential gain from adopting a cognitive-neuroscience perspective for understanding one critical applied issue: motivation. We will leave it to subsequent theorists to elaborate the social extensions of this system. Our focus is more limited, namely, to show that neuropsychologically grounded theories of cognitive processes can be used to develop a better understanding of how human information processing capacities constrain motivation and decision-making. Specifically, we analyze the information processing mechanisms that may underlie one popular optimization theory – VIE theory (Vroom, 1964) – from a neural network perspective. We believe the resulting framework provides a clearer picture of when and how people are able to arrive at subjectively optimal decisions. Though our intent is not to dwell on the social extensions of this perspective, we should point out that optimality in decision-making (or lack of it) has long been at the heart of many macro-level theories. It is the foundation stone of neoclassical economic theory (Simon, 1991), a central aspect of transaction cost economics (Williamson, 1975), and the focus of bounded rationality theories of organizational decision-making (Cyert & March, 1963; March & Simon, 1958).

VALENCE, INSTRUMENTALITY, EXPECTANCY THEORY AND MOTIVATION

Expectancy theory and Vroom’s (1964) ValenceInstrumentality-Expectancy (VIE) version of this theory, in particular, has motivated substantial research activity in the applied literature on motivation.

22 Lord, Hanges, and Godfrey According to VIE theory, people’s actions and choices are lawfully related to the preferences and affective reactions they have for certain outcomes (i.e., valences), their beliefs about whether certain actions lead to particular outcomes or performance levels (i.e., expectancies), and their perception of the association between primary and secondary outcomes (i.e., instrumentalities). People are assumed to consciously review their valences, instrumentalities, and expectancies when deciding on a particular course of action, and it is believed that they will act in a subjectively optimal manner once these factors have been considered. VIE theory captured researchers’ attention for two reasons. First, its emphasis on internal thought processes rather than external reinforcement schedules was consistent with the evolving zeitgeist of the 1960s. Second, researchers were intrigued by the explicit hypotheses provided in VIE theory about the way people combine their valences, instrumentalities, and expectancies to produce motivated behaviour. Specifically, Vroom postulated that people undergo a process that was summarized in two mathematical equations. The first equation states that the preference for a primary outcome (i.e., the outcome that directly follows from an action) is a function of the perceived connection between that primary outcome and other possible outcomes (referred to as secondary outcomes in the VIE literature) and the attractiveness or valence of these secondary outcomes. More specifically: (1) where Vj, represents the valence of a particular primary outcome, Ijk represents the perceived or “cognized” instrumentality connecting the primary outcome to any secondary outcomes, and Vk represents the attractiveness of the various secondary outcomes. The second equation states that the motivational force for a particular action or choice is a function of the perceived valence of the primary outcomes and the expectancy that the action would lead to these various outcomes. More concisely:

(2) where Forcei represents the motivational force driving someone to choose a particular action, Eij represents the expectancy or “momentary belief” that the particular action leads to a particular primary outcome, and Vj represents the valence of a particular primary outcome.

A substantial amount of research activity was stimulated by Vroom’s ideas, and the influence of this theory extended beyond the motivation literature into many areas of decision-making. Indeed, VIE theory influenced scholars working on such diverse topics as leadership (e.g., House, 1971), compensation (e.g., Lawler, 1971), and turnover (e.g., Summers & Hendrix, 1991). Seminal reviews of the VIE research all noted some empirical support, but have also identified several problems. These problems have ranged from mismeasurement of the three key constructs (e.g., Ilgen, Nebeker, & Pritchard, 1981; Schwab, Olian-Gottlieb, & Heneman, 1979) to the use of an inappropriate experimental design (i.e., between-subjects design) to validate a within-subjects theory (e.g., Mitchell, 1974). A recent meta-analysis by Van Eerde and Thierry (1996) found that despite these problems, there is support for the theory. More specifically, Van Eerde and Thierry report that valence, instrumentality, and expectancy by themselves, as well as their combined VIE products, predicted performance, effort, intentions, preferences, and choice with average correlations between .33 and .65. Despite this long history and empirical support, a comparison of VIE theory (or other types of subjective expected utility models) with current knowledge about human information-processing capabilities suggests that the theory contains several untenable assumptions. First, such optimization models assume that people consider a wide variety of actions, outcomes, and instrumentalities when making choices. Unfortunately, people often lack the needed information, have trouble accessing a large set of information from memory, or use information in a suboptimal manner. Simon (1945) offered an alternative model that assumed that people only consider a limited amount of information, and stop considering alternatives once a satisfactory outcome has been identified. Others (e.g., Lord & Maher, 1990; Nisbett & Ross, 1980) suggested that more heuristic models based on either simplified computational rules or extensive feedback were more reasonable descriptions of how people typically make choices In addition to the extensive informational requirements, the mental computational requirements directly implied by the two VIE equations are unrealistic. It is now recognized that humans are relatively slow, serial processors of symbolic-level information. Therefore, conscious application of even relatively simple mathematical rules takes considerable time. For example, mental arithmetic research has shown that although addition and subtraction of single digits can be performed in about 200 ms, the multiplication of a pair of elements (as implied in VIE theory, i.e., E1V1) takes

Rethinking VIE Theory 23

Figure 1. Feedforward multilayer neural network.

approximately 2,000 ms. A realistic decision situation might consist of 10 alternatives mapping onto 10 firstand 10 second-level outcomes. Following Vroom’s model, decision-making in such a situation would require thousands of separate multiplication steps and take several hours of a person’s full processing capacity. Although Vroom’s theory may also describe the effects of implicit cognitive operations rather than conscious computations, precisely how implicit processes can lawfully combine valences, expectancies, and instrumentalities in an optimal manner has not been specified. In his introduction to the Classic Edition of his 1964 book, Vroom (1995) himself notes that although he still believes that much behaviour can be thought of as intentionally rational or goal directed, “The level of processing required by expectancy theory is rarely possible and would represent one extreme, found mainly on relatively simple choice problems where the alternatives are clear and information is readily available” (p. xix). At the other extreme, behaviour is more under the control of emotions or habit. Vroom goes on to note that the perceived valences of outcomes wax and wane and may vary with sequential attention to goals, thus valences may lack properties such as transitivity in choices that would be expected of utilities.

In sum, there is an inconsistency or paradox in this literature that needs to be resolved. Namely, how can the empirical support for VIE theory or other optimization models be explained, given the logical inconsistencies between the tenets of this theory and what we now know about the information-processing limits of individuals? We believe that this inconsistency is due to the VIE theory emphasis on volitional, explicit processes in decision-making. This paradox can be resolved by reframing VIE theory in terms of a neural network information-processing model. Recent cognitive science research suggests that neural networks operating in parallel can perform many computational activities, and may be capable of performing the VIE equation computations automatically and rapidly, assuming that one has extensive experience with the problem situation. In the present paper, we reconceptualize the VIE model by developing a neural network model of choice and motivated behaviour. Basically, our network model reframes VIE theory from a model implying conscious consideration of expectancies and instrumentalities to a model involving more automatic, implicit use of these decision elements. Before discussing the specifics of our reconceived VIE theory, we will provide a general overview of neural networks.

24 Lord, Hanges, and Godfrey Basics of Neural Networks Work on neural network models was started in the 1940s to illustrate how relatively simple computational processes could produce complex perceptual and cognitive behaviours (Hanges, Lord, Godfrey, & Raver, 2001). Over the years, the computational capabilities of neural networks have become more complex and researchers have used these networks for two different purposes. One way that researchers use neural networks is to conduct statistical analyses. This use of neural networks emphasizes the computational capabilities of these networks and it is suggested that these networks might be better at uncovering nonlinear relationships than more traditional statistical techniques (e.g., Collins & Clark, 1993; Garson, 1998; Hanges et al., 2001; Somers, 1999). The second way that researchers use neural networks is to focus on their theory building contribution by developing models of various phenomena. For example, neural networks have been developed for many types of cognitive and social processes such as grapheme-to-phoneme conversion in speech coding (Clark, 2001), categorization of stimuli and extraction of category prototypes (Smith, 1996; Smith & DeCoster, 1998), causal attributions (Van Overwalle, 1998), and leadership perceptions (Lord, Brown, & Harvey, 2001; Lord, Brown, Harvey, & Hall, 2001), to provide just a few examples. One particularly important aspect of neural networks is their capacity to exhibit contextual sensitivity at a fine-grained level (Clark, 2001), as reflected in theories of cultural influences on behaviour (Hanges, Lord, & Dickson, 2000; Strauss & Quinn, 1997). It should be realized that even though this second approach to neural networks is more conceptual, the computations underlying how neural networks process information is the same as for the statistical use of neural networks. In the present paper, we rely on both the conceptual and computational implications of neural networks to illustrate how this approach can be used to reconceptualize VIE theory in more implicit, subsymbolic terms. Figure 1 shows a simplified neural network. The building blocks of neural networks are “nodes” that integrate information from either the outside environment or from other nodes in the network. Usually, researchers make a distinction between three types of nodes: input, hidden, and output. The network shown in Figure 1 is composed of three input nodes (I1, I2, and I3), three hidden nodes (H1, H2, and H3), and one output node ( O 1). The input nodes bring information about the outside environment into the network. The hidden nodes combine the information from the environment as well as other hidden nodes. Each connection arrow is associated with a weight

indicating that the input nodes are more strongly associated with some hidden nodes than others. In particular, the hidden node combines the information from the input nodes in the following manner:

(3) where I1 , I2 , and I3 represent the information provided by the three input nodes, w1 , w2 , and w3 are the connection weights associated with each of these input nodes, b is a constant, called the bias, that adjusts the sensitivity of each hidden node, and H1 represents the activity level1 of the first hidden node in Figure 1. In this example, H1 and H2 are basically combining information from the environment, whereas H 3 processes the information from H 1 and H 2. Although there are two hidden layers shown in Figure 1, networks can be designed with one or more hidden layers. The information continues to be processed throughout the network until it comes to the output nodes. These output nodes produce some action, thereby connecting the activity of the network to the outside world. The network shown in Figure 1 is called a multilayered, feedforward network. The network is considered feedforward because the neuron activation flows unidirectionally from the input variables through the hidden units to the output variables (Bechtel & Abrahamsen, 1991). It is called a multilayered network because there are hidden layers between the input and output layers. This network builds its knowledge about the environment by learning associations among the nodes (Smith, 1996). Specifically, as learning occurs, the weights connecting units in the networks change (Bechtel & Abrahamsen, 1991). For example, the Hebbian learning rule is one algorithm used by computational researchers to train a neural network to its environment. It is based on the proposal of Donald Hebb (1949) in his classic book, The Organization of Behavior, that when two neurons in the brain were both active, the strength of their connection would be increased. This learning rule changes the connection weights as a function of the covariation between the activity of the network’s components during each encounter with the environment. If the activation levels of the two components are similar (e.g., both are 1 For the purposes of this article, we are assuming that the activation level of a particular hidden node is direct function of the linear combination of connection weights and input node activity. It should be noted that networks with nonlinear relationships between node activity and input information could be designed.

Rethinking VIE Theory 25

Figure 2. A recurrent neural network.

either positive or negative), the weight connecting these two components is increased. On the other hand, if the activation levels of the two components are different, the weight connecting these two components is reduced. In this way, the network “learns” about associations among the components of the network and eventually, after repeatedly encountering a particular pattern of input over time, a stable pattern of activity emerges within the network. This stable pattern can be thought of as an internal representation of the statistical regularities in an external environment. As such it reflects an implicit mental model of frequently encountered environments. Johnson-Laird (1983, 1989) has persuasively argued that responses to task environments are not direct, but rather are mediated by internal mental models that develop from perceptions. Importantly, he stresses that it is the relations among components in mental models that are critical to developing meaning or for reasoning based on mental models. Although Johnson-Laird focuses on symbolic systems, we believe the same principles could apply to the neural networks, but that the nature of meaning or reasoning is qualitatively different, implicit rather than explicit, and may be heavily dependent on the affective consequences of networks.

Researchers working with neural networks tend to call these patterns of stable activation schemas rather than mental models. These schematic patterns of activity in the network become increasingly efficient (i.e., require less effort to evoke) as the number of encounters with the original input pattern increases. Eventually, these patterns are so efficient that they are automatically activated (i.e., chronically accessible) even when encountering input stimuli that do not exactly match the input stimuli originally learned (Bechtel & Abrahamsen, 1991). In neural network models, schemas are not conceptualized as fixed structures residing in memory. Rather, they are continually regenerated over time. Schemas are seen as being sensitive to situational differences because of differences in the input information. These context differences can have subtle or significant effects on the output or behaviours of people (Strauss & Quinn, 1997). Thus, these models easily allow for the context-specific nature of schemas. In other words, people may behave differently in different situations, but that does not mean that their schemas have changed permanently. Rather, it means that the schema, when regenerated, differed from previous situations because of differences in the environment of the current situation.

26 Lord, Hanges, and Godfrey TABLE 1 Key Benchmark Principles and Their Operationalization in Neural Network Models of VIE Theory Benchmark Principle

Operationalization in Neural Network Model

1. Neural networks are computational systems.

Feedforward networks perform matrix multiplication (i.e., compute the inner product of vectors) (Churchland & Sejnowski, 1992). That is, the mathematics implied by such a network is that pairs of components from two vectors (i.e., vector of connection weights and vector of input activity) are multiplied and these products are summed.

2. VIE concepts are represented both implicitly and explicitly.

Decision-making components can be represented as values associated with particular nodes (i.e., Valences) or as learned weights connecting two nodes (i.e., Instrumentalities or Expectancies).

3. Outcome patterns are Gestalts, not sets of discrete symbols.

The parallel flow of activation from input to output units create meaningful patterns such as expectancy weighted outcome “bundles” associated with different actions.

4. Patterns organize into meaningful structure over time.

Networks learn prototypical activation patterns among nodes from repeatedly encountered inputs, even though no single exemplar is contained in the entire pattern (Smith & De Coster, 1998).

5. Networks compute vector similarity or degree of pattern matching.

If second-level valences are represented as a prototype for “good outcomes,” then networks create goodness judgment for expectancy weighted outcome patterns based on the similarity of activated pattern to stored prototypes.

6. Goodness of fit could be used to select actions in a flexible manner.

Alternative set could be sequentially evaluated with optimization occurring if entire set was evaluated or satisficing occurring if process terminated with first alternative to create “acceptable” fit.

Another type of neural network is shown in Figure 2. This neural network is called an interactive or recurrent network. This network is more complex than the multilayered feedforward network and several differences between these two networks can be immediately spotted by comparing Figures 1 and 2. First, in the recurrent network, all the variables are connected with each other. Second, recurrent networks allow bidirectional activation of neurons throughout the network (Bechtel & Abrahamsen, 1991). The behaviour of a neural network is mainly a function of its architecture. Therefore, the different designs of these two networks imply that they behave differently. In particular, the feedforward networks are useful for understanding how information from the outside world interact with internal processes to evoke either attributions or behaviour (e.g., leadership attributions or follower behaviour). Recurrent networks, on the other hand, have been used to model pattern recognition aspects of memory and information processing (e.g., prototype matching processes) (Smith, 1996; Smith & DeCoster, 1998). This necessarily brief introduction to neural networks barely scratches the surface of this issue, and researchers interested in learning more about neural networks are referred to Bechtel and Abrahamsen (1991), Churchland and Sejnowski (1992), Hanges et al. (2001), and Garson (1998). We now turn our attention to VIE theory and use these two neural networks to develop an alternative view of VIE processes.

Resolving the Paradox of VIE theory We have previously noted that meta-analytic studies (Van Eerde & Thierry, 1996) find considerable support for VIE theory, particularly when within-subject designs are used. Yet it is also clear that the informational and processing demands of formal VIE models and most other optimization models are inconsistent with what we know about how people operate to solve complex, symbolic problems. As indicated earlier, we believe that this paradox can be resolved by positing that the computations implied by optimization models are actually performed sub symbolically (i.e., automatically) by neural networks. In this section, we examine this proposition more closely, investigating the potential for feedforward and recurrent networks to mimic the computational operations suggested by VIE theory. We also consider ways that VIE theory and general conceptualizations of rationality might be reconceived to make them more compatible with the functioning of neural networks. Table 1 summarizes the main steps in our reasoning, providing benchmarks and key discoveries we have made while developing this argument. Our initial realization that neural networks were appropriate computational devices for implementing procedures implied by expectancy valence theories came from Churchland and Sejnowski (1992). As we note in the first row in Table 1, these authors emphasized the connection between the matrix computations emphasized in the statistical approach to neural networks and the

Rethinking VIE Theory 27

Figure 3. A feedforward multiplayer model of VIE theory.

conceptual application of these models. In other words, Churchland and Sejnowski explained that feedforward networks are essentially devices for performing computational operations on information. The operation of the network is the same regardless of whether the information is perceptual or mathematical. Multiplication operations are analogous to the flow of activation along paths. At the end of a path, the resulting activation depends on the strength of initial input activation multiplied by the weight or strength of connection between input and output nodes. Summation processes occur within an output node as activation from all connected upstream sources is aggregate. Our second insight concerned the way we conceptualize the expectancy and instrumentality components in our neural network model. We believe that although the alternative actions and outcomes considered by an individual are nodes in a neural network, it is best to think of expectancies and instrumentalities as connection weights in these networks which through learning come to reflect regularities in task environments. Figure 3 is a pictorial representation of our initial neural network reformulation of VIE theory. As can be seen in this figure, we conceptualized the

alternative actions considered by a person as initial sources of activation. This is equivalent to positing that networks are activated when an individual thinks of or consciously evaluates an action. This conscious evaluation can be triggered by the external environment, as shown in Figure 3, or by one’s internal goals. When a person thinks of one possible action (e.g., Possible Action 1 in Figure 3), a series of outcomes that are likely to result from that action receive activation, whereas outcomes that are less likely to occur would receive less activation or may even be inhibited. The connection between first- and second-level outcomes would also reflect weights that determine the flow of activation between these units. These weights correspond to instrumentalities in VIE theory. We assume that weights are learned based on action/outcome or outcome/outcome associations in the experienced task environments. Since the same learning rule would govern the learning of both expectancies and instrumentalities, they can potentially have the same range (-1 to 1), with negative numbers indicating degree of inhibitions and positive numbers the degree of activation. Thus, this formulation of expectancies differs slightly from Vroom’s view that they reflect subjective probability estimates. The conceptualization of instrumentalities and

28 Lord, Hanges, and Godfrey

Figure 4.

A hybrid feedforward, recurrent network model of alternative choice.

expectancies as connection weights has several advantages. First, if one considers the Hebbian learning rule discussed earlier, the connection weights shown in the Figure 3 network are revised based on the degree of association between the various outcomes and/or actions. Over time, one would expect high weights to reflect strong associations between alternative actions and outcomes. This increase (diminishment) of connection weights is precisely what is believed to happen with expectancies as a person gains experience. Second, conceptualizing instrumentalities and expectancies as connection weights suggests that these constructs really reside in a person’s implicit rather than explicit knowledge. Implicit knowledge guides our behaviour, but it cannot be formally articulated. This may be one reason for some of the measurement problems that have limited tests of optimization models such as VIE theory, since subjects are typically asked to explicitly estimate expectancies and instrumentalities rather than merely use them in a natural environment. Finally, as highlighted by Equation 1, conceptualizing expectancies and instrumentalities as connection weights leads one to the conclusion that it is possible for a model to be consistent with the current understanding of implicit information processing capabilities of humans and still pos-

tulate that the integration of the valence, instrumentality, and expectancy components to reach a decision occurs in a multiplicative fashion. Although the multiplicative integration of information is relatively consistent with VIE theory, it should be stressed that our neural network model suggests that this integration occurs outside of conscious awareness (i.e., subsymbolic level of processing). Evaluation of a potential action may then flow automatically from conscious attention to that alternative. That is, when a potential alternative action is in working memory (for whatever reason), it is in a state of heightened activation, and thereby may serve as an input activation source for related constructs. Similarly, thinking of a particular first-level outcome may then activate an array of second-level outcomes in proportion to their learned instrumentalities. The operation of such a process begins with conscious attention, which creates activation that spreads through learned networks automatically and in parallel, much like dropping a rock in a pool of water starts with a discrete action but then the flow of associated energy ripples outward in an automatic manner. Ponds, however, have an amorphous structure, whereas neural networks have a discrete structure based on prior learning. Moreover, this learning translates the

Rethinking VIE Theory 29 structure of task environments into the structure of decision-relevant schema. Symbolic models of problem-solving also recognize that complexity (Simon, 1978) and rationality (Anderson, 1991) of human behaviour stem from the structure of human task environments. We are suggesting here that much of this structure is learned and used implicitly. We began with this conceptualization of an alternative evaluation process, which is nicely modeled by a simple feedforward network. One need only posit that alternatives are considered sequentially to see how each alternative could, in turn, create a unique pattern of outcome activation, which, in total, was analogous to the motivational force or attractiveness of that action. However, by itself, such a network does not perform precisely the computation implied by expectancy valence theory. Rather than summing the products of expectancies and valences as implied by Equation 1, the feedforward network shown in Figure 3 sums the products of information from alternative action states to the primary and then secondary outcomes. We thought of two ways to address this inconsistency between VIE theory and our neural network model. One way would be to conceptualize the network as operating in reverse, with outcomes being the initial source of activation and the action being the downstream node that summed expectancy weighted valences. Although this reversal would be consistent with the information flow suggested by the VIE equations, such a model did not seem very plausible in a psychological sense. We thought a better approach was to live with the notion that considering an alternative activated a pattern of outcomes through a feedforward network, and that this pattern, as a whole, was a psychologically meaningful construct. Outcomes would then operate as a Gestalt rather than as discrete symbols that are processed separately and in isolation. This, then, is our third key insight as shown in Table 1. The notion that decision-makers automatically process patterns of outcomes in parallel has considerable appeal from a neural network perspective. This is essentially what we do when perceiving objects as being meaningful; bundles of features are automatically mapped into appropriate concepts or meanings. For example, objects with legs, a flat surface, and a back are perceived as chairs, or people with specific sets of features are recognized as belonging to ethnic-, racial-, gender-, or trait-based categories. Moreover, neural networks are thought to be particularly good at modeling such processes (Smith, 1996; Smith & DeCoster, 1998). Such perceptual or categorization processes require two components – an initial input pattern created by a feedforward network and a

learned schema or prototype. These two components are shown in the left and right halves of Figure 4. One may ask whether it makes sense to think of rational decision-making as involving a comparison of an expected pattern to a schema or prototype. Two lines of reasoning suggest that it does. First, models of schema development suggest that prototypes are built up from repeatedly encountered patterns (Sherman, 1996). Second, recent theories of decision-making often posit a comparison of an action to images of what we are trying to achieve (Beach, 1990). Interestingly, recurrent neural networks can assess the similarity of two vector patterns, which we can conceptualize as an alternative/outcome vector and a prototype or image vector (Churchland & Sejnowski, 1992). This process can work for both unfamiliar and familiar decision situations, although our focus is on the latter. For unfamiliar decision problems, people may explore a number of alternative actions. These patterns can then be abstracted on the fly to form a prototype or a definition of a good outcome, even though no specific pattern precisely matches this prototype. (See Smith and De Coster, 1998, for an illustration of now this can be done using recurrent neural networks.) For familiar decision problems, decisionmakers may sequentially explore a number of alternatives, which, in turn, activate outcome patterns. Each pattern can be compared to a learned prototype to evaluate its goodness of fit or optimality. For both unfamiliar and familiar situations, the evaluation process for a specific alternative involves a comparison of its associated outcomes with a learned prototype. This process is described on the fourth and fifth rows in Table 1 and is shown in Figure 4. In other words, alternatives could be recognized as being good outcomes just as people are recognized as being extroverts rather than introverts, or buildings are recognized as being houses rather than stores. Thus, what we are suggesting is that the decision processes inherent in motivational processes may be closer to a perceptual than an explicit mathematical process. However, the perceptual processes, if implemented by the type of neural network processes we have suggested, could do a fairly good job of mimicking explicit optimization processes. Moreover, they can do this very quickly (typically less than one second) and with little apparent effort. What then is the role for conscious processing in such a motivational or decision process? First, as already mentioned, it may reflect the initial input to neural networks in which alternatives are activated by conscious visualization processes. Second, it may help one store perceptions in memory so that sequentially evaluated alternatives can be compared. Such con-

30 Lord, Hanges, and Godfrey scious processing can also provide a degree of flexibility. If one stopped at the first good alternative, one essentially would have a satisficing process (March & Simon, 1958). If one retained only the best alternative in memory, but considered an entire set of alternatives, the process would be closer to optimization. In sum, then, the six principles shown in Table 1 suggest one way in which neural networks can be of value, helping to explain how slow, serial symbolic processors can perform very well in familiar decision environments by blending symbolic-level operations with subsymbolic processes. This idea obviously is preliminary and needs conceptual refinement and empirical examination. Such development may come in the form of simulations to explore some of the key aspects of the suggested model, as we show in a latter part of this paper. In addition, at some stage of investigation, one would ideally compare the functioning of a decision simulator to the fine-grained behaviour of actual decision subjects. Other Evidence for Dual-Process Models of Behaviour Recent theory and empirical work suggest that such a hybrid symbolic/subsymbolic model of motivation is quite reasonable. For example, Schneider and Oliver (1991) simulate human information processing in terms of a combination of symbolic and neural network processes (CAPS2), maintaining that these two architectures are complementary. Smith and DeCoster (2000) review several extensive areas of social and cognitive psychology research, showing that each is consistent with a dual-process model in which people use both a slow learning memory system located in the neocortex (i.e., neural network) and a fast learning, symbolic system that is mediated by the hippocampal memory system. As McClelland, McNaughton, and O’Reilly (1995) show, the slow learning system extracts statistical regularities from one’s environment, which are reflected in the changes in weights in a neural network-based system, and it is the basis for implicit learning. The fast learning hippocampal system protects prior learning from interference while at the same time allowing the rapid formation of associations to create new patterns. Applying this theory to our VIE example, we are simply maintaining that the slow learning system can use statistical regularities (i.e., weights) extracted from familiar environments to optimize choices even though the fast learning, conscious system cannot readily perform the computational processes required for optimization. The potential value of these two memory systems in guiding human behaviour can be seen in another of their interactive features. Although neural networks

store statistical regularities in neocortical systems, they may do this in a compressed form that losses noncritical details. However, when these structures are retrieved into hippocampal systems, patterns may be created subject to constraints from one’s current context. This rather abstract idea has important implications, for it suggests that schemas or prototypes can be contextualized to a current context. For example, social categories like one’s idea of a good leader may depend on context (see Hanges et al., 2000; Lord et al., 2001). Similarly, decision processes may be optimized in a contextually sensitive manner when the outcome prototypes that are reinstated in hippocampal memory systems are adjusted to reflect contextual constraints. For example, choice of a place to dine may be contexutalized by morphing one’s generic dining schema into a schema that is more compatible with current constraints (time, money, social companions, etc.). Though based on the functioning of actual physical structures, this process is reminiscent of Simon’s notion of bounded rationality in that only a limited, contextually determined definition of optimality is used in decision-making. Other Decision-Making Applications Bayesian decision-making shares with subjective expected utility theories like VIE theory the notion that there are mathematically optimal ways to make decisions. In conceptual terms, Bayesian theory says that the conditional probability of an event occurring depends on the product of an event’s base rate times a likelihood ratio involving the conditional information. People, however, tend to ignore base rates when problems are presented in symbolic form. For example, in making medical diagnosis, people tend to focus on the fit of a pattern of symptoms with a salient disease, while underusing information about the probability of a particular disease occurring. However, this underuse of base-rate information disappears when decision-makers actually experience base rates rather than receiving such information in symbolic form (Christensen-Szalanski & Beach, 1982; ChristensenSzalanski & Bushyhead, 1981). This finding implies that the statistical base-rate information gained implicitly through experience is somehow more useful than equivalent information provided in symbolic form. One possibility, consistent with our VIE theory example, is simply that implicit computational processes gained through experience can be used more easily than explicit computational processes. This may be because experienced information is implicitly incorporated into the weights of neural networks, allowing the use of dual-processes to make Bayesian decisions.

Rethinking VIE Theory 31 TABLE 2 Simulation Parameters __________________________________________________________________________________________________ Outcomes _______________________________________________________________ Restaurants First-order Second-order __________________________________________________________________________________________________ 1. Guzzato’s Tasty Food Healthy Food 2. Enchanted Asian Pleasant Atmosphere Good Conversation 3. Dino’s Diner Good Price Show off to Others 4. Meatball Warehouse Please Client Blow Budget 5. Taste of Italy MSG Headache Smokey Dining Room 6. Szechwan Salon Make Business Deal __________________________________________________________________________________________________

Role of Experience One critical aspect of our suggested decision-making model, which is also critical in understanding the under use of base-rate information in Bayesian decision-making, is the role of experience. In our proposed model, we chose to represent expectancy and instrumentality information in terms of weights that are learned through experience. Thus, expectancy and instrumentality information is implicit rather than explicit. By this we mean that as one gains experience, weights are automatically modified to reflect the statistical association between inputs and outputs. Further, the effects of weights are implicit, being reflected in the flow of activation. Similarly, we conceptualized the learning of an outcome prototype as being an implicit process that is based on experience. There are several important predictions that follow from this conceptualization of expectancies. First, the development of optimum decisions requires experience in choosing and implementing alternatives. As we have already shown in the area of Bayesian decisionmaking, people are better at extracting base-rate information from experience than they are at using it when supplied in a formal sense. Our prediction, then, is that decision optimization should follow a learning curve that is similar to weight modification in feedforward networks with Hebbian learning and prototype learning in recurrent networks Second, we would expect that asking subjects to provide expectancy and instrumentality information would not be a very good means to measure weights. One would need to develop more implicit, data-driven means to estimate weights. For example, one might explore a policy-capturing approach where one estimates the association between dimensions of alternatives (which could be conceptualized as hidden units in a three-layer feedforward network) and resulting outcomes. Similarly, asking subjects to describe prototypical outcome patterns might not be as effective as estimating prototypes using data-driven implicit processes such as the alternative comparison processes that provide inputs to procedures like Pathfinder.

Such procedures begin by developing a proximity matrix among alternative outcomes from a series of similarity judgments or card-sorting procedures. This input is then used to construct representations of subjects’ mental models based on estimations of underlying networks. Viability of Neural Network Model: A Simulation Complex models, such as the one we have developed, often do not operate as intended because of problems with internal logic or because they are poor descriptions of human behaviour. The issue of internal logic of a complex model can be addressed by using computer simulations. Such simulations can demonstrate that the proposed model can be implemented and work as expected. We investigated this internal logic issue by simulating the model we described earlier. We also expanded this model to incorporate the effects of current goals as a means to show how dualprocess models of behaviour could operate. To test this model, we wrote a MATLAB (Version 6.1) computer program that followed the general decision model shown in Figure 4. Specifically, our program simulated the cognitive processes of an individual deciding between six different restaurants. We assumed that the individual evaluated each restaurant on the basis of five first-order outcomes and six second-order outcomes. These first- and second-order outcomes were chosen because they could be clustered under four superordinate goals (i.e., health goals, business goals, cost-related goals, social goals). Table 2 shows the name and characteristics of the six restaurants, the first-order outcomes, and the secondorder outcomes. The connection weights for our simulation were generated from the authors’ collective impressions of six specific restaurants.2 For each restaurant, we estimated the probability that dining at that restaurant would lead to each first-order outcome (i.e., expectan2 The names of the actual restaurants that we used to develop our simulation have been changed in Table 2.

32 Lord, Hanges, and Godfrey

Figure 5.

Flow diagram illustrating how the simulated neural network processed information.

cies). A positive connection weight indicated increasing likelihood that choosing to eat at a particular restaurant would result in a particular first-order outcome. A negative connection weight indicated an increasing likelihood that choosing to eat at a particular restaurant would not result in obtaining the particular first-order outcome. A connection weight of 0.0 indicates no connection between choosing a particular restaurant and the likelihood of obtaining a particular first-order outcome. Following this general “policy capturing” approach, we also estimated the connection weights (i.e., valences) between the first-order and second-order outcomes. All connection weights ranged from -1.0 to + 1.0. An understanding of how the simulation worked can be obtained by examining Figure 5. The top portion of this figure is a flow diagram that shows how the activation of the network starts with some external event (i.e., left side of figure) that makes the individual think about one restaurant (e.g., someone suggests Guzzato’s as a possible place to eat). In the simulation, the activation level provided by the external environment is represented by ao vector. The w10 vector represents how accurately the simulated individual interprets the external event. In all of the present

simulations, the simulated individual was believed to have an almost perfect interpretation of the external environment. Once the person starts to think about a particular restaurant (i.e., a 1 vector), the activation starts to spread to the first-order (a2 = a1 * w21 + b2 + random error) and then to the second-order (a3 = a2 * w32 + b3 + random error) outcomes (i.e., activation spreads from left to right of figure). However, this flow of activation does not only flow in a single direction. Rather, as the person thinks about the restaurant, the activation level reverberates through the network. The activated second-order outcomes can affect the activation levels of the first-order outcomes (a2 = a3 * w23 + b2 + random error). In other words, the activation spreads from right to left in Figure 5. After some set time, the simulated individual reaches a conclusion about the restaurant (i.e., some stable level of activation is achieved). This stable activation pattern is called the restaurant prototype. In our simulation, the restaurant prototype is obtained after the simulated individual thinks about the restaurant for 20 iterations. We designed our simulation so that the external event (i.e., left to right flow) had a moderately stronger influence on the outcome activation than did the

Rethinking VIE Theory 33

Figure 6. Exemplars of two restaurant prototypes.

34 Lord, Hanges, and Godfrey TABLE 3 Restaurant Choices ______________________________________________________________________________________________________________________________________________________________ Ideal Standards ______________________________________________________________________________________________________________________________________________________________ Cost = 1.0; Business = 0.0; Cost = 0.5; Business = 0.0; Cost = -.3; Business = 1.0; Social = 0.0; Health = 0.0 Social = 0.2; Health = 1.0 Social = 0.1; Health = -.3 ______________________________________________________________________________________________________________________________________________________________ • Szechwan Salon (SSD=1.35) • Enchanted Asian (SSD=8.19) • Guzzato’s (SSE=2.00) • Taste of Italy (SSD=1.44) • Meatball Warehouse (SSD=1.50) ______________________________________________________________________________________________________________________________________________________________ • Dino’s Diner (SSD=5.68) • Dino’s Diner (SSD=9.07) • Enchanted Asian (SSE=6.45) ______________________________________________________________________________________________________________________________________________________________ •______________________________________________________________________________________________________________________________________________________________ Enchanted Asian (SSD=17.84) • Guzzato’s (SSD=9.21) • Dino’s Diner (SSE=19.07) • Guzzato’s (SSD=24.52)

• Taste of Italy (SSD=9.68)

• Taste of Italy (SSE=23.15) • Szechwan Salon (SSE=23.58) • Meatball Warehouse (SSE=24.24) ______________________________________________________________________________________________________________________________________________________________ • Szechwan Salon (SSD=10.28) • Meatball Warehouse (SSD=10.40) ______________________________________________________________________________________________________________________________________________________________ Note: Smaller values of SSD indicate greater fit between ideal standard and restaurant prototype.

internal (i.e., right to left) activation flow. Finally, before estimating each restaurant prototype, we reset the network’s nodes to zero (i.e., no activation level). We did this to eliminate any carryover effects when estimating multiple restaurant prototypes. Hypothesis 1: Different restaurant prototypes (i.e., activation patterns) will be obtained for the six restaurants.

As discussed earlier, we believe that the person decides to go to the restaurant whose prototype most closely matches the person’s ideal standard. The flow diagram in the bottom portion of Figure 5 shows how we generated this ideal standard. Specifically, we created a hierarchy among the four goals by creating a unique activation pattern (i.e., the g vector in Figure 5) for the goals. In a completely separate process, this g vector influenced the activation level of the firstorder outcomes (aideal2 = g * w2g + b2 + error) and second-order outcomes (aideal3 = g * w3g + b3 + error). The activation process reverberated through the system until a stable activation pattern was achieved. The sum of the squared difference (SSD) between a restaurant’s prototype and the ideal standard was computed. The restaurant whose prototype most closely matches the ideal standard (i.e., the restaurant with the smallest SSD) is the restaurant that is chosen. One primary difference between our neural network model and the original Vroom conceptualization is that we maintain that the choice of restaurant will differ, even when the connection weights (i.e., expectancies and valences) are held constant, as a function of the particular ideal standard that is activated. To demonstrate whether this is a reasonable state-

ment, we developed three different ideal standards. For the first ideal standard, the cost goal was the most important (i.e., activation of cost goal = 1.0, activation of all other goals = 0.0). For the second ideal standard, the health goal was the most important (i.e., activation = 1.0), cost was of moderate concern (i.e., activation = 0.5), social goal was of minor concern (i.e., activation = 0.2), and business goals were not important (i.e., activation = 0.0). For the third ideal standard, the business goal was the most important (i.e., activation = 1.0), social goals were of minor concern (i.e., activation = 0.2), the cost goal actually had a negative salience (i.e., activation = -.3) because the dinner was assumed to be paid via a business expense account. The health goal also had a slightly negative salience (i.e., activation = -.3). We expect that the choice of restaurant will differ depending upon which ideal standard was activated. Hypothesis 2: The choice of restaurant will vary as a function of the saliency of the four superordinate goals.

We ran a total of 100 independent replications for each of the three ideal standards. These replications enabled statistical testing of the average SSDs for the six restaurants. RESULTS

The first hypothesis predicted that the simulation would produce different prototypes for the six restaurants even though the prototypes were affected by internal influences, external influences, and random error. This hypothesis was supported. The top panel in Figure 6 shows the average prototype for Guzzato’s

Rethinking VIE Theory 35 restaurant. Some of the outcomes were positively activated (e.g., “Good Food,” “Atmosphere,” “Please Client”) whereas other outcomes were deactivated (e.g., “Good Price” and “Smokey Dining Room”). The bottom panel in Figure 6 shows the average prototype for Dino’s Diner. In contrast to Guzzato’s restaurant, only one outcome was strongly activated (i.e., “Smokey Dining Room”) for Dino’s Diner. Further, the “Good Price” outcome was moderately activated for this restaurant and all of the other outcomes were deactivated for Dino’s Diner. Thus, consistent with the first hypothesis, different prototypes emerged for the six different restaurants. The second hypothesis predicted that different choices would be made depending on the ideal standard that is chosen. As Table 3 indicates, there was overall support for this hypothesis. Each column represents the fit of the restaurant prototypes for a particular ideal standard. Greater fit between the restaurant prototypes and the ideal standard is reflected by smaller restaurant SSD. For each ideal standard considered (i.e., for each column of Table 3), we used the Scheffe’s post-hoc test to assess whether the restaurants significantly differed in terms of their fit (SSD) with the ideal standard. Restaurants in the same cell do not significantly differ from each other. Thus, for the first ideal standard in which cost is completely driving the decision (i.e., left column of Table 3), it appears that three restaurants were equally likely to be chosen (i.e., Szechwan Salon, Taste of Italy, Meatball Warehouse) but for the second ideal standard (i.e., middle column), the Enchanted Asian restaurant was the clear winner. Consistent with our second hypothesis, different restaurants were chosen as a function of the three ideal standards. DISCUSSION

This simulation demonstrates that the conceptual model outlined in our paper behaves in a fashion consistent with our predictions. As we discussed previously, the simulation has demonstrated that multiple prototypes can be stored and extracted from a multilayered, feedforward network after its connection weights (i.e., expectancies and valences) have been specified. The simulation also illustrates that a network designed to perform a relatively cognitively simple comparison process can produce some rather complex choice behaviour. In addition, this simple cognitive comparison model is based on an integration of information concerning the valance, instrumentality, and expectancy of outcomes. Finally, the simulation showed that our network model will yield different predictions even when the connection weights are held constant by simply changing the

saliency of the different goals determining the ideal comparison standard. Thus, the model provides one example of how dual-processing models could operate. Although the simulation has demonstrated that the proposed model can operate as hypothesized, it is important to realize that the simulation did not provide any empirical support for the appropriateness of the model. Future studies need to test whether this proposed model predicts choice behaviour better than the original VIE theory predictions. For example, a possible study could start off following the traditional VIE theory experiment in which participants provide ratings of valences, instrumentalities, and expectancies for a set of outcomes. Following the procedure we used in our simulation, the VIE information can be used as connection weights to create different network model specific for each participant in the experiment. The accuracy of the predictions from these individual-specific networks can be assessed under a number of experimentally created conditions in which the participants’ goal hierarchies are varied. This type of research will be very helpful in assessing the descriptive accuracy of the proposed model. So What? A legitimate question that can be asked at this point is “ So What? What is the value of this complex model for understanding motivation or decision-making?” This question can be answered on three levels. First, does the model describe people’s behaviour? As we just noted, that is still an open question. Second, does the model provide new theory or integrate old theories? Here we are on firmer ground. We reviewed evidence advocating dual-processing theories of human behaviour, and in our simulations showed how such a theory can operate by looking at the interaction of our neural network model with currently operating goals. Simply put, goals affect the choices that are made, and our approach can model this interaction of conscious and automatic processes. There are many other potential ways that conscious processes can interact with the operation of the choice networks we have examined. Although goals are normally thought of as conscious mechanisms that regulate human behaviour (Latham & Locke, 1991), other less conscious processes can also operate to change the functioning of the type of choice network we have proposed. One is context or immediate history. We are currently working on incorporating how the context affects the decision processes. Early results indicating that the prototypes for choices considered later in the decision-making process are affected by the prototypes previously generated. Another

36 Lord, Hanges, and Godfrey relevant factor is framing, a topic that has dominate the decision-making literature for the past 25 years (Kahneman & Tversky, 1979). We are currently examining whether gain versus loss frames can operate in a manner analogous to goals to alter outcome preferences. A third factor is emotions. We are also examining models of emotions which operate through the biasing factor (b) in Equation 3. This factor can be thought of as the ease with which relevant units can fire, a construct that is roughly analogous to the effects of emotions on neurotransmitters. Our point in mentioning these extensions is to illustrate the flexibility of the general model we have developed. By combining the interaction of conscious processes with neural networks of the type illustrated in this paper, we can develop more coherent “dual-process” models of how choice and motivation operate. Future research can compare these models to the behaviour of real people in real decision contexts. Our third response to the “So What?” question centres on the value of such models for guiding decision making theories and applications. The appreciation of human information processing limitations in the last half of the 20th century has directed theorists and practitioners away from rational models; focusing instead on human weakness that result is use of heuristics, framing, intuition, or emotions to guide choice. Our model suggests an alternative perspective by showing that in familiar environments people can behave in “rational” ways using automatic processes based on neural networks. Further, we have shown how dual-processes approaches can easily adapt these automatic network processes to current goals. Thus, we suggest that people can be both rational and flexible. However, this capacity comes from the subtle interplay between conscious and unconscious processes involved in choice, not from sole reliance on conscious computational processes. This is a new view of how effective decision-making operates that may have many practical implications. CONCLUSION

In this paper we have shown how models of information processing inspired by neuropsychological methods can offer fresh insight into the potential operation of familiar models of motivation and decisionmaking. Although interesting, we believe that this is merely a tip-of-the-iceberg phenomena in terms of the potential relevance of neuropsychological theory and findings to applied psychology. For example, emerging research is linking data based on neuroimaging technology with functional models of behaviour to gain a greater understanding of the structure of affect and its role in motivation (Carver, 2001). Similarly,

fundamental differences in personality (extroversion and neuroticism) have been shown to predict how extensively human brain locations react to positive or aversive stimuli, respectively (Canli et al., 2001). Also, expertise in a subject area may have a specific neurological correlate as activation of the lateral fusiform gyrus may be a critical marker of visual expertise and recognition of familiar but not novel faces (Ochsner & Lieberman, 2001). Such effects may have profound implication for advances in applied psychology. Thus we conclude where we began, with the recognition that emerging new areas such as cognitive neuroscience or social cognitive neuroscience reflect a new source of theoretical perspectives and methodological techniques that could be of immense value to applied psychologists. We would like to thank Todd Troyer for his advice and guidance with earlier versions of the computer simulations reported in this paper. Address correspondence to Robert G. Lord, Department of Psychology, University of Akron, Akron, OH 44325-4301. (Phone: 330-972-7018; Fax: 330-973-5174; E-mail: [email protected]).

Résumé Ce manuscrit utilise une reformulation du modèle théorique de Vroom (1964) (la théorie VIE) pour illustrer la valeur potentielle des modèles de processus cognitifs axés sur la neuropsychologie. La théorie de Vroom suppose que les décisions des personnes sont déterminées par leurs réactions affectives à certains résultats (valences), à certaines croyances quant à la relation entre les actions et les résultats (attentes) et les perceptions de l’association entre les résultats primaires et secondaires (instrumentalités). L’une des principales critiques de ce type de théorie demeure que les calculs nécessaires prennent un temps irréaliste et dépassent souvent la capacité de la mémoire de travail. Dans cet article, nous soutenons que si un individu éprouve une situation problème prolongée, il peut traiter les décisions concernant cette situation à l’aide des réseaux neuraux qui fonctionnent implicitement de manière à ce que les ressources ne s’épuisent pas simplement par les calculs; plutôt, les calculs sont exécutés implicitement par les réseaux neuraux. En pensant à la théorie VIE dans la perspective du réseau neural, au moins un des problèmes s’élimine et plusieurs autres nouvelles introspectives dans la prise de décision se découvrent. Nous nous servons de la méthodologie de la simulation pour montrer qu’un tel modèle est tout autant viable et il peut refléter les effets des objectifs courants sur les processus de choix.

Rethinking VIE Theory 37

References Abdi, H., Valentin, D., & Edelman, B. (1999). Neural networks. In M. S. Lewis-Beck (Ed.), Quantitative applications in the social sciences. Thousand Oaks, CA: Sage Publications. Beach, L. R. (1998). Image theory: Theoretical and empirical foundations. Mahwah, NJ: Lawrence Erlbaum Associates. Bechtel, W., & Abrahamsen, A. (1991). Connectionism and the mind. Cambridge, MA: Blackwell Publishers. Canli, T., Zhao, Z., Kang, E., & Gross, J. (2001). An fMRI study of personality influences on brain reactivity to emotional stimuli. Behavioral Neuroscience, 1, 33-42. Carver, C. A. (2001). Affect and the functional bases of behavior: On the dimensional structure of affective experience. Personality and Social Psychology Review, 5, 345-356. Chirstensen-Szalanski, J. J., & Beach, L. R. (1982). Experience and the base-rate fallacy. Organizational Behavior and Human Performance, 29, 270-279. Christensen-Szalanski, J. J., & Bushyhead, J. B. (1981). Physians’ use of probabilistic information in a real clinical setting. Journal of Experimental Psychology: Human Perception and Performance, 7, 928-935. Churchland, P. S., & Sejnowski, T. J. (1992). The computational brain. Cambridge, MA: MIT Press. Clark, A. (2001). Mindware: An introduction to the philosophy of cognitive science. New York: Oxford University Press. Collins, J. M., & Clark, M. R. (1993). An application of the theory of neural computation to the prediction of workplace behavior: An illustration and assessment of network analysis. Personnel Psychology, 46, 503-524. Cyert, R. M., & March, J. G. (1963). A behavioral theory of the firm. Englewood Cliffs, NJ: Prentice Hall. Garson, G. D. (1998). Neural networks: An introductory guide for social scientists. Thousand Oaks, CA: Sage Publications. Hanges, P. J., Lord, R. G., & Dickson, M. W. (2000). An information-processing perspective on leadership and culture: A case for connectionist architecture. Applied Psychology: An International Review, 49, 133-161. Hanges, P. J., Lord, R. G., Godfrey, E., & Raver, J. L. (2001). Modeling nonlinear relationships: Neural networks and catastrophe analysis. In S. G. Rogelberg (Ed.), Handbook of research methods in industrial and organizational psychology. Hebb, D. O. (1949). The organization of behavior. New York: Wiley. House, R. J. (1971). A path-goal theory of leader effectiveness. Administrative Science Quarterly, 16, 321-338. Ilgen, D. R., Barnes-Farrell, J. L., & McKellin, D. B. (1993). Performance appraisal process research in the 1980s: What has it contributed to appraisals in use?

Organizational Behavior and Human Decision Processes, 54, 321-368. Ilgen, D. R., Nebeker, D. M., & Pritchard, R. D. (1981). Expecrancy theory measures: An empirical comparison in an experimental simulation. Organizational Behavior and Human Decision Processes, 28, 189-223. Johnson-Laird, P. N. (1983). Mental models. Cambridge, MA: Harvard University Press. Johnson-Laird, P. N. (1989). Mental models. In M. I. Posner (Ed.), Foundations of cognitive science. Cambridge, MA: MIT Press. Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47, 263-291. Latham, G. P., & Locke, E. A. (1991). Self-regulation through goal setting. Organziational Behavior and Human Decision Processes, 50, 212-247. Lawler, E. E. (1971). Pay and organizational effectiveness: A psychological view. New York: McGraw-Hill. Lord, R. G., Brown, D. J., & Harvey, J. L. (2001). System constraints on leadership perceptions, behavior, and influence: An example of connectionist level processes. In M. A. Hogg & R. S. Tindale (Eds.), Blackwell handbook of social psychology: Group processes. Lord, R. G., Brown, D. J., Harvey, J. L., & Hall, R. J. (2001). Contextual constraints on prototype generation and their multilevel consequences for leadership perceptions. The Leadership Quarterly, 12, 311-338. Lord, R. G., Foti, R. J., & De Vader, C. (1984). A test of leadership categorization theory: Internal structure, information processing, and leadership perceptions. Organizational Behavior and Human Performance, 34, 343378. Lord, R. G., & Maher, K. J. (1990). Alternative information-processing models and their implications for theory, research, and practice. Academy of Management Review, 15, 9-28. March, J. G., & Simon, H. A. (1958). Organizations. New York: John Wiley. McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Why there are complementary learning systems in the hipppocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory. Psychological Review, 102, 419457. Mitchell, T. R. (1974). Expectancy models of job satisfaction, occupational perference and effort: A theoretical, methodological, and empirical appraisal. Psychological Bulletin, 81, 1053-1075. Nisbett, R., & Ross, L. (1980). Human inference: Strategies and shortcomings of social judgment. Englewood Cliffs, NJ: Prentice-Hall. Ochsner, K. N., Lieberman, M. D. (2001). The emergence of social cognitive neuroscience. American Psychologist,

38 Lord, Hanges, and Godfrey 56, 717-734. Schneider, W., & Oliver, W. L. (1991). An instructable connectionist/control architecture: Using rule-based instructions to accomplish connectionist learning in a human time scale. In K. VanLehn (Ed.),. Architectures for intelligence (pp. 113-145). Hillsdale, NJ: Lawrence Erlbaum. Schwab, D. P., Olian-Gottlieb, J. D., & Heneman, H. G. (1979). Between-subjects expectancy theory research: A statistical review of studies predicting effort and performance. Psychological Bulletin, 86, 139-147. Sherman, J. W. (1996). Development and mental representation of stereotypes. Journal of Personality and Social Psychology, 70, 1126-1141. Simon, H. A. (1945). Administrative behavior: A study of decision-making processes in administrative organization. New York: Free Press. Simon, H. A. (1978). The science of the artificial (2nd ed.). Cambridge, MA: MIT Press. Simon, H. A. (1991). Cognitive architectures and rational analysis: Comment. In K. VanLehn (Ed.), Architectures for intelligence (pp. 25-39). Hillsdale, NJ: Lawrence Erlbaum. Smith, E. R. (1996). What do connectionism and social psychology offer each other? Journal of Personality and Social Psychology, 70, 893-912. Smith, E. R., & DeCoster, J. (1998). Person perception and stereotyping: Simulation using distributed representations in a recurrent connectionist network. In S.

J. Read & L. C. Miller (Eds.), Connectionist models of social reasoning and social behavior (pp.111-140). Mahwah, NJ: Lawrence Erlbaum. Smith, E. R., & DeCoster, J. (2000). Dual-process models in social and cognitive psychology: Conceptual integration and links to underlying memory systems. Personality and Social Psychology Review, 4, 108-131. Somers, M. J. (1999). Application of two neural network paradigms to the study of voluntary employee turnover. Journal of Applied Psychology, 84, 177-185. Strauss, C., & Quinn, N. (1997). A cognitive theory of cultural meaning. New York: Cambridge University Press. Summers, T. P., & Hendrix, W. H. (1991). Development of a turnover model that incorporates a matrix measure of valence-instrumentality-expectancy perceptions. Journal of Business and Psychology, 6, 227-245. Van Eerde, W., & Thierry, H. (1996). Vroom’s expectancy models and work-related criteria: A meta-analysis. Journal of Applied Psychology, 81, 575-586. Van Overwalle, F. (1998). Causal explanation as constraint satisfaction: A critique and a feedforward connectionist alternative. Journal of Personality and Social Psychology, 74, 312-328. Vroom, V. H. (1964). Work and motivation. New York: Wiley. Vroom, V. H. (1995). Work and motivation. San Francisco, CA: Jossey-Bass. Williamson, O. E. (1975). Markets and hierarchies: Analysis and antitrust implications. New York: Free Press.