David Henderson and Terence Horgan

 

This is the penultimate draft of a paper forthcoming in Karsten Steuber (ed.), Empathy and Agency: The Problem of Understanding in the Social Sciences, Westview Press (forthcoming).

 

I. Naturalized Epistemology

            Epistemology has recently come to more and more take the articulate form of an investigation into how we do, and perhaps might better, manage the cognitive chores of producing, modifying, and generally maintaining belief-sets with a view to having a true and systematic understanding of the world.  While this approach has continuities with earlier philosophy, it admittedly makes a departure from the tradition of epistemology as first philosophy.  Such investigations are termed “naturalized epistemologies” insofar as they help themselves to scientific information regarding human cognitive capacities, and regarding the possibilities for social organization in the pursuit of the classical epistemic goal.  Epistemology on this model has been on the rise at least since Quine’s classic manifesto, “Epistemology Naturalized,” declared our principled freedom from the obligation to provide a first philosophy.

            So, epistemology (naturalized) is concerned with human capacities useful in the pursuit of that classical epistemic end of a true and systematic understanding of the world.  When it yields a model of how we do commonly manage this pursuit with whatever success we now enjoy, it provides a descriptive account of our epistemic competence.  Even such a descriptive account has a significant normative force, as it codifies what is acceptably effective in present practices, counciling us away from the residue of cognitive performances that do not so contribute.  When the epistemological account employs information regarding the bases and plasticity of our present cognitive capacities in order to construct a model of how we might yet better manage our pursuit of our epistemic goal, we produce meliorative models of epistemic competence (Henderson 1994a, Kitcher 1992).

            Taken together, the capacities constituting our (descriptive or meliorative) epistemic competence will have certain characteristics making for their (joint) effectiveness in the pursuit of our epistemic goal.  One very prominent general characteristic is the most familiar of externalist virtues: the reliability of processes implementing the capacities.  Here, famously, reliability is a matter of the tendency of those processes to produce true beliefs in the environment that the epistemic agent inhabits as a matter of fact.  However, this virtue cannot be the whole story.  For one thing, it is a matter of the ratio of true beliefs to the total set of individual beliefs produced when capacities are set to work in the agent’s environment—and thus a set of capacities could be reliable and yet not tend to the production of sets of beliefs that have the systematicity desired.[1]  Our epistemic goal is systematic true understanding—not some loose assortment of individual true beliefs.  So, truth-conducivity (reliability) is not the sole or decisive epistemic virtue.

However, truth-conducivity and systematicity-conducivity do not jointly exhaust the externalist virtues to be recognized.  Suppose that, by dumb luck, some agents live an epistemically sheltered life—living in an narrow environment where it is particularly easy to formulate systems of true beliefs.  For example, in the possible world in which these hypothetical agents live, populations are so homogeneous that no sensitivity to sample representativeness is needed—the most hasty generalization from the most accidentally drawn sample will produce a true generalization in their world.  Perhaps the agents are such damn-the-torpedos generalizers, and thus employ processes that are reliable (and even conducive to systematicity) in their world of wimpy epistemic demands.  Are their processes to be commended?  It seems to us that the answer is not an unqualified “yes.”  In view of the exigencies and uncertainties that are characteristic of epistemic life, it is desirable for agents to employ processes that are reliable (and conducive to systematicity) in a significant range of epistemically possible worlds—including those in which populations may be heterogeneous, for example.  It is probable that an agent employing such processes is employing processes that are reliable and system-conducive, whatever the epistemically possible world that agent occupies (as long as it is not a perversely difficult one).  Horgan and Henderson (MS) call the indicated epistemic virtue robustness.

So, on the naturalized epistemological approach adopted here, cognitive capacities are epistemically valuable insofar as they take their place in a model of our descriptive or meliorative competence.  For a capacity to have a place in such a model is a matter of the tendencies arising from its combination with select other capacties that we have or could have (given the base of capacities on which creatures like us build).  For a capacity to have a place in our models of epistemic competence is for it to be a component of a set of capacities that is conducive to the truth and systematicity of our belief systems, and is so in a way that is not precarious.

In this paper, we provide a rough characterization of the role of simulation in our epistemic competence for understanding others.  Our characterization is intended to apply both to our descriptive competence and to our meliorative competence, as it reflects the limits and general promise of both theory-based and simulation-based understanding.  We later suggest some implications for the social sciences.  However, here at the beginning, we insist that nothing we say here is to minimize the importance of the sort of sustained interpretive back-and-forth that is involved in constructing a body of information regarding someone’s belief-system.  In keeping with the hermeneutic circle, the general understanding of someone’s belief-system results as a systematic accretion from  increasingly successful treatments of particular cases; where these in turn depend upon an increasingly comprehensive general understanding of our subject(s).  That is, a general understanding of our subjects’ belief-and-desire system serves to inform our understanding of particular cases, while our faith in the general picture depends on apparent successes in the particular cases.  The focus in this paper is on the capacities that are brought to the task of generating our understandings of particular cases.  We believe that both the application of theory and simulation play roles in the treatments of particular cases, and that this is compatible with an appropriate treatment of the extended task of developing an interpretive scheme for a people (Henderson 1993, 1994b).

Although we conceive of this paper as a piece of naturalized epistemology, we do not here draw explicitly on many particular empirical works.  We work at one remove, drawing instead on a general understanding of human cognitive processing that is elsewhere developed at greater length, and with more direct comment on empirical work in cognitive science (Horgan and Tienson, 1996).  We sketch this background understanding in section III.

II. Clarifying the Question, “What is the Epistemic Role of Simulation?”

            The simulation of another agent, in the sense of interest here, is a matter of taking certain cognitive processes “off-line,” and setting them to work on input that is similar to those that the other (supposedly) had or will have; this is undertaken so that the resulting processing  parallels what has or will take place in the other.  It is a matter of what Goldman (1989) called “process-driven” simulation in which the agent’s own cognitive processes, acting on “pretend input,” generate (“pretend”-) beliefs, desires, or intentions as outputs.  This is contrasted with “theory-driven” simulation in which some representation of the interaction of parameterized variables within a system is put to work on some hypothetical values—as in computer simulations of a system.  In process-driven simulation, the simulating agent need have no representation of the transitions to which the simulated agent is disposed.  For, if the simulation is successful, the simulator’s own capacities for dealing with inputs of the sort that the other had are themselves similar to the other’s, and thus the transitions carried out in the simulator’s own off-line capacities parallel those in the simulated.

In its rudiments, simulation may be thought of as a three-stage affair.  First, there is the generation of a set of “pretend-beliefs,” “pretend-desires,” “pretend-perceptions,” and perhaps other input.  Here ‘pretend’ is employed to signal a certain sort of discontinuity.  These states are generated in ways that may be discontinuous with the common processes of belief and desire generation.  They are generated so as to be highly similar to those that the subject of the simulation had in the relevant episode.[2]  Further, the capacities into which these are to be fed will themselves be operating “off-line,” which is to say in a way that is discontinuous with their normal operation.  After all, here, the system will generate as output “beliefs” and “desires” that are not themselves integrated into that agent’s own belief- and desire-sets, and it will generate “intentions” that do not eventuate in actions.  Second, there is the off-line processing itself, in which the simulating agent’s own cognitive capacities are put to work on the input.  While these have been taken off-line, they supposedly work pretty much as they would otherwise.  Third, there is the ascription to the other of the beliefs, desires, intentions, and reasoning instanced in the simulation.

It is important to note what is not entailed by the above rudimentary understanding of simulation.  Centrally, there is no suggestion that simulation works alone when it is called into play.  Both proponents and opponents of simulation-theory have sometimes written as if simulation and discursive information were like oil and water—supposing that they do not mix.  Accordingly, one would need to get by with simulation, untutored by discursive information, as some simulation-theorists would seem to have it, or with discursive representations unsupplemented by simulation, as some theory-theorists would have it.  Perhaps these ways of seeing the matter are a hangover from the old debate between methodological separatists and naturalists.  Proponents of simulation theory have sometimes sought to formulate their position so as to minimize opportunities for informing simulation by discursive information—perhaps fearing that such information (as rudimentary theory) would represent the nose of a camel that they seek to corral in a pen devoted to the natural and biological sciences.  On the other hand, opponents have been happy with this framing of the matter.  It would seem to make simulation “cognitively inpenetrable”—that is, insensitive to discursive information that one might have.  Since such inpenetrability  is not observed, opponents happily conclude that simulation is not epistemically important.  The above rudimentary understanding of simulation does not encourage such moves.

Our characterization allows simulation to be informed by discursive information both in the generation of input and in the ascription to another of parallels (in contents as well as kind) to the off-line processes that are then spun out.  What is required for simulation—what is at its core—is a set of cognitive capacities on which the simulator can draw in order to put into play off-line processes that are similar to the simulated agent’s.  It is this use of the simulator’s own cognitive capacities that distinguishes simulation from those cognitive processes that turn on manipulation of representations of that agent’s processes.   So, the core of the simulation is free of theory representing the agent’s cognitive capacities and the inputs on which they work.[3]  But, this leaves much scope for discursive information to play a role in determining what input to employ, what cognitive processes to call into play, and what confidence to place in the judgment that resulting process actually paralleled processes in the agent simulated.

The insistence on unsullied simulation or unsupplemented theory seems farfetched, from a general epistemological point of view.  Our competence in dealing with a subject matter commonly comprises a mix of strategies, jointly providing multiple modes of epistemic access to matters of interest.  The access provided by one method may complement that provided by another.  One may compensate where the other encounters difficulty.  And, where multiple strategies are applicable, they may check up on each other and provide opportunities for feedback and refinement.  Henderson (1995) argues that simulation-employing and theory-employing approaches to understanding and explaining others complement each other in these ways.  The present paper continues this line of thought.

The place for simulation within our epistemic competence will then turn on how it works in yoke with other capacities that we have or might have.  The issue that needs to be addressed is whether the capacity for simulation, in combination with other capacities, has the epistemic virtues of reliability, robustness, and systematicity—and, if so, just what roles are assigned to simulation in the package (or packages) of capacities with these virtues.

 

III.        One Role for Simulation.

            We here begin to explore cooperative roles for our capacity for simulation.  Some recent work by Jane Heal (1996) provides one piece of the puzzle.  Heal insists that we do have a good deal of discursive information regarding persons, their psychological states, and their cognitive tendencies.  She allows that this information may itself be sufficiently extensive and interrelated to be called a “theory.”  However, she insists that our theory of mind, such as it is and is likely to be, is not sufficient to explain our ability to understand and explain others.  She argues that, at the level detailing the interaction of contentful states, at just the level that theory would need to be applicable in order to yield the sort of concrete predictions and explanations for which adult humans have a significant facility, our psychological theory must be incomplete.  It may help us along, but, at some point, we must turn the particulars over to simulation.  The theory may, presumably, have been important in framing the simulation, but simulation is needed to really deal with tracing out the particulars of the processing to be understood.

            Heal, as we understand her, argues for two key claims.  First, no theory of intentional states and their interactions could be as successfully predictive of people’s actual psychological state-transitions as we humans typically are.  Second, even if there were such a predictive theory, it still would be extremely unlikely that humans employ such a theory in understanding one another.  The first claim is more fundamental; for, if no suitably predictive theory over intentional states exists at all, then people’s predictive successes in understanding one another could not possibly be based entirely upon their use of such a theory.

            Heal defends the first claim by appeal to four important facts about thinking. First is the enormity of the information which any normal adult human agent has acquired.  The amount and diversity of information contained in any normal human’s belief system is impressive, even awe inspiring.   Second, the epistemic status of a thought is holistic.  That is, it depends on the rest of the agent’s beliefs.  (Or, in the case of thoughts that are desirings, its appropriateness depends on the rest of the agent’s beliefs and desires.)  The holistic character of epistemic justification has at least two prominent aspects.  The first aspect is that the epistemic status of a given belief depends on the set of relevant beliefs that the agent has.  But, any one belief is potentially relevant to any other beliefs, with this being a function of what further beliefs the agent possesses.  So, the set of beliefs relevant to a given belief, and thus that belief’s epistemic status, depends on the agent’s global belief-set.  The second aspect is that the epistemic status of a given belief is, in some measure, dependent on global characteristics of the containing set of beliefs—characteristics such as simplicity, or overall explanatory power.  The first aspect is particularly important for us here.  It means that the epistemic agent must somehow be sensitive to what other beliefs within that agent’s enormous set of beliefs are relevant to a given candidate belief, and that the agent should then bring these relevant beliefs together in maintaining or rejecting the candidate belief.

            The third fact that Heal emphasizes is that, remarkably, human cognitive agents are fairly good at just this.  They show remarkable sensitivity to what are relevant beliefs, even in unusual cases where normally unrelated beliefs become relevant.  Her fourth point is that, to understand and account for human cognition, we must be able to deal with people’s sensitivity to selective holistic relevance, and we must be able to track this capacity to cope with far-flung beliefs even in cases where the constellation of relevant beliefs turns out to be unusual.

            Heal designates a theory that could model human cognition with these characteristics, and with the nuanced twists and turns that result, a theory of relevance.  In light of the four facts about human thinking lately mentioned, she maintains, it is very unlikely there can be a general and systematic theory of relevance.  That is, there can be no psychological theory, stated in terms of intentional states and their interactions, that is as successfully predictive of people’s psychological state-transitions as humans actually are.

            Heal also argues that even if there were such a theory of relevance, such a theory would be unlikely actually to be employed by humans as a means of predicting and understanding one another.  She offers two considerations in defense of this claim.  First, it is somewhat implausible that we have such a powerful implicit theory without this being reflected in an ability to make some significant start in its partial articulation—an ability that is strikingly lacking.  Second, and more importantly, we should be wary of the suggestion that humans could implicitly deploy such a theory in understanding one another, given the very enormity of the task it is supposed to accomplish for us.  It is supposed to provide the basis for theoretically modeling the ways that information gets accommodated as human cognizers continuously update their sets of beliefs.  It must describe how agents commonly hit upon the contents of their belief-sets that need to be considered in accommodating a new piece of information—drawing a frame around those that are relevant and are to be put into the mix.  It needs then to say how the resulting sets of beliefs will interact to produce a new ensemble.  Given the size and complexity of human belief-systems, given the nuanced and novel ways that pieces may come to fit together, actually applying such a theory would be a daunting task indeed.  “As an information storage and processing task and given the range of our actual psychological competence,” she says,  “dealing with this imagined theory is orders of magnitude more formidable than dealing with any other tacit theory that has been proposed, e.g., for grammar or folk physics” (p. 84).[4]

            Heal observes that that the inability to uncover a theory of relevance is at the core of a problem that researchers in AI have persistently run into, the so-called frame problem, which Jerry Fodor has usefully characterized as “the problem of putting a ‘frame’ around the set of beliefs that may need to be revised in light of specified newly available information” (Fodor 1983, 112-13).   Heal’s discussion suggests (although she does not quite say this explicitly) that the computational theory of mind, which construes mental processing as the rule-governed manipulation of syntactically structured mental representations on the basis of their syntactic structure, is itself mistaken in the absence of a theory of relevance.

            We are very sympathetic to Heal’s arguments. Concerning her principal claim that a general, systematic, theory of relevance is not possible, similar pessimistic considerations about the likely non-existence of a theory of relevance, and about how this makes the frame problem an apparently in-principle problem that seriously threatens the very possibility of accounting for cognitive processes like belief-formation within the computational theory of mind, are set forth by Fodor (1983).  He stresses (i) the same holistic aspects of non-demonstrative inference emphasized by Heal, (ii) the key role of these holistic aspects in the persistent failure of attempts to produce a rigorous theory of non-demonstrative inference, and (iii) the parallel between the seemingly in-principle failure to systematize non-demonstrative inference and the apparently in-principle unsolvability of the frame problem within computational cognitive science.  Such considerations are further elaborated by Horgan and Tienson (1996), who conclude both (i) that there cannot be a predictive psychological theory, couched in terms of intentional psychological states, of the kind that Heal designates a theory of relevance, and (ii) that the computational conception of human mental processing is very likely mistaken.[5] (See also Henderson and Horgan, forthcoming.)

            Cognitive science evidently needs to replace the computational conception of mind with something more powerful.  An alternative, more powerful, picture of human cognition has indeed begun to emerge from within cognitive science.  It draws upon connectionist modeling, and also on a form of mathematics that is natural for describing connectionist models--dynamical systems theory.  This nonclassical framework for cognitive science is described at length in Horgan and Tienson (1996).  Here we offer a very brief summary, with emphasis on features that are especially germane to the debate about how humans understand one another.

            Connectionism emerged as a large-scale research program in the 1980's, largely in response to the recurrent, recalcitrant, difficulties that led Fodor to his bleak conclusions about the prospects for classical AI modeling central cognitive processes.  A connectionist system, or neural network, is a structure of simple neuron-like processors called nodes or units.  Each node has directed connections to other nodes, so that the nodes send and receive excitatory and inhibitory signals to and from one another.  The total input to a node determines its state of activation.  When a node is on, it sends out signals to the nodes to which it has output connections, with the intensity of a signal depending upon both (i) the activation level of the sending node and (ii) the strength or "weight" of the connection between it and the receiving node.  Typically at each moment during processing, many nodes are simultaneously sending signals to others.

            When neural networks are employed for information processing, certain nodes are designated "input" units and "output" units, and potential patterns of activation across them are assigned interpretations.  (The remaining nodes are called "hidden units.")  Typically a "problem" is posed to a network by activating a pattern in the input nodes; then the various nodes in the system simultaneously send and receive signals repeatedly until the system settles into a stable configuration; the semantic interpretation of the resulting pattern in the output nodes is its "answer" to the problem.[6]

            The most striking difference between such networks and conventional computers is the lack of an executive component.  In a conventional computer the behavior of the whole system is controlled at the central processing unit (CPU) by a stored program.  A connectionist system lacks both a CPU and a stored program.  Nevertheless, often in a connectionist system certain activation patterns over sets of hidden units can be interpreted as internal representations with interesting content, and often the system also can be interpreted as embodying, in its weighted connections, information that gets automatically accommodated during processing without getting explicitly represented via activation patterns.

            Connectionist models in cognitive science have yielded particularly encouraging results for cognitive processes like learning, pattern recognition, and so-called multiple-soft-constraint satisfaction (i.e., solving a problem governed by several constraints, where an optimal solution may require violating some constraints in order to satisfy others--e.g., successfully classifying a given three-legged animal as a dog, even though dogs have four legs).

         In a connectionist system, information is actively represented as a pattern of activation.  When the information is not in use, that pattern is nowhere present in the system; it is not stored as a data structure.  The only representations ever present are the active ones.  On the other hand, information can be said to be tacitly present in a connectionist system--or "in the weights," as connectionists like to say--if the weighted connections subserve representation-level dispositions that are appropriate to that information.  In the terminology of Horgan and Tienson (1995, 1996), such information constitutes morphological content in the system (i.e., content  that is present in virtue of the system's structure), rather than explicitly-represented content.  Among the apparent advantages of connectionist systems, by contrast with classical computational systems, is that morphological information "in the weights" gets accommodated automatically during processing, without any need for a central processing unit to find and fetch task-relevant information from some separate memory banks where it gets stored in explicit form while not in use.

        Learning is conceived quite differently within connectionism than it is within classicism, since connectionist systems do not store representations.  Because learning involves the system's undergoing weight changes that render its representation‑forming dispositions appropriate to the content of what is learned, learning is the acquisition, "in the weights," of new morphological content.

            The apparent moral of frame-type problems in computational cognitive science is that human cognitive transitions are evidently too subtle and too complex to conform to any general and systematic theory of relevance, and hence are too subtle and too complex to be subserved by computation over representations.  Connectionism provides a promising alternative paradigm.  Horgan and Tienson’s (1994, 1996) proposed non-classical foundational framework is inspired partly by the emergence of the connectionist movement, partly by some particularly suggestive connectionist work such as Berg (1992), Pollack (1990), and Smolensky (1990, in press), partly by frame-type problems facing classicism, and partly by the natural links between neural networks and the branch of mathematics called dynamical systems theory.  The mathematics of dynamical systems is more powerful than the discrete mathematics of algorithms; in effect, algorithmic systems are a limited special case of dynamical systems.[7]

            Horgan and Tienson call this non-classical approach the dynamical cognition framework (for short, the DC framework), because of the central role it assigns to dynamical systems theory.  A key feature of the DC framework is that much of the information that gets accommodated in cognitive processing is morphological, not occurrent.  Rather than being explicitly represented via occurrent states in the course of processing, such information is instead accommodated implicitly "in the weights."[8]  Morphological content is especially important with respect to the holistic aspects of cognitive processing.  According to the DC framework, these aspects are not subserved by processes that update beliefs (and other informational states) by computationally manipulating explicit representations of all epistemically relevant information (thereby implementing a general, systematic, and precise theory of relevance).  The moral of frame-type problems in classical cognitive science is that this is just not possible.   Rather, the holistic aspects of cognitive tasks like belief fixation are primarily subserved morphologically: cognitive transitions are automatically appropriate to large amounts of implicit information and to holistic normative-justificatory relations involving that information.  According to the picture of cognitive processing provided by the DC framework, holistic aspects of cognitive processing must be subserved morphologically; this is an essential aspect of Nature's "design solution" to the problem of avoiding frame-like breakdowns in human cognition, and to the closely related problem of managing significant inductive reasoning without intractability.

            The DC framework does not deny the possibility of systematic psychological generalizations, or the claim that such generalizations have an important explanatory role in cognitive science.  But the kinds of generalizations the framework envisions will not, in general, generate predictions of people’s specific psychological state-transitions; the generalizations will not be as predictively powerful as we are ourselves in our ability to anticipate and understand one another.  There are at least two reasons why the generalizations will fall short of such predictiveness.  First, normally they will be soft (in the terminology of Horgan and Tienson 1990, 1996); i.e., they will have ineliminable ceteris paribus clauses that advert to potential psychology-level exceptions—and not merely to non-psychological exceptions like physical malfunction (e.g., having a stroke) or external interference (e.g., being hit by a bus).

            A second reason why the kinds of generalizations allowed for by the DC framework normally will not be strongly predictive involves likely features of such generalizations that Heal herself emphasizes.  (Her focus is on generalizations that are part of our common-sense knowledge of persons, but the point carries over to certain generalizations that cognitive science might produce within the DC framework.)  Concerning the kinds of theoretical knowledge that we already possess, she says:

We are capable of stating explicitly a fair amount of about the sort of beings we take people to be, the factors which influence them and how they interrelate.  For example they can perceive what is in spatial proximity to them through their various senses and can remember past events and can in these ways acquire beliefs.  They have desires and, under the guidance of their beliefs, form projects on how to fulfil them.  They feel emotions which are liable to influence their patterns of reasoning.  And so on.... Such generalities say nothing directly on beliefs or projects about particular subject matters, e.g. under what circumstances a doctor will believe that a patient has measles or when a restaurant customer will order a meal.  (1996, pp. 78-9)

Since the kinds of generalizations Heal cites in this passage are about psychological state-types as broad classes, in general they do not combine with specific facts about a person’s total psychological state at a given moment, to yield predictions about the psychological state-transitions the person will undergo thereafter (not even hedged predictions with an attached ceteris paribus qualifier).  This is because such generalizations do not have specific instantiations of the right form—viz., the form of a conditional statement whose antecedent describes a combination of specific beliefs, desires, projects, etc. possessed by someone at a specific moment in time, and whose consequent describes such a combination of specific psychological states at a subsequent moment.  As one might put it, they are not plug-in generalizations—the idea being that one cannot take a description of a person’s total psychological state at an initial moment, and then plug it into an instantiation of the generalization, in order to obtain a prediction of the person’s total psychological state at a later moment.

            Let us take stock. Heal is not doubting that theory and theoretical modeling may play some role in how we manage to understand others, and neither are we.   In what follows, we attempt to say more about the complementary roles of theory and simulation.  For now, the point is just that the application of theory alone, either explicit or implicit, cannot plausibly be the sole component of our competence here.  In general, the theoretical generalizations available to us will fail to be strongly predictive, because normally they will soft, and/or they will not be plug-in generalizations.  Even allowing that discursive information regarding cognitive tendencies can affect how we understanding others, still, at some point, the task of tracing out how certain contentful states would interact with the agent’s many others will commonly need to be turned over to simulation.  Here, the one seeking to understand relies on his or her own processes to trace out just where a set of beliefs would lead.  In Heal’s terms, since we lack a theory of relevance, in understanding others, and in “grappling with content,” we ultimately need to be able to simulate.

IV. The Dynamic Duo

We have argued that simulation will need to play an important role in our understanding others.  So, although our understanding may be shaped by discursive information in ways yet to be sketched, commonly, at some point, we will need to rely on simulation to trace out the global interaction of content.  To further appreciate the role for simulation in understanding others, we will need to get clear on the roles also played by processes driven by discursive information.  How does our capacity for simulation hook up with other capacities to make for our full epistemic competence in understanding others?

Neither simulation nor theory-application threaten to monopolize the capacity for understanding others, and they should not be seen as competing models of our competence, for neither alone is up to satisfying the market for understanding.  The reason is simple: each, with a minimum of supplementation, has significant blind spots.  Since their blind spots are not identical, we do well epistemically, to employ simulation and theoretical modeling as complementary processes—and to employ hybrid processes.  For purposes of this paper, we may do well to recast considerations found in Henderson (1995).  To begin, let us think in terms of the two extremes within the possible mixes of theory and simulation.  This will allow us to take some stock of the blind spots of each.

First, consider the extreme of reliance on theory, where discursive information (some sufficiently developed to be called “theory”) carries the maximum load in understanding.  To imagine a relatively pure case of theory-reliance, suppose that we have a powerful psychological theory—implicit or explicit—that characterizes rather many of the sorts of transitions between contentful states to which humans are disposed.  We may also suppose that we can draw upon an extensive discursive characterization of what certain folk believe and desire.  Finally, we might add a characterization of how perceptual beliefs are likely to arise in folk against a background of beliefs.[9]

Now, even supposing this highly favorable theoretical situation, we should acknowledge that the strategic use of off-line processes plays a limited but crucial role in applying all this discursive information.  Again, the theory characterizes the transitions to which agents are disposed in terms of relations between contentful states—typically not mentioning particular beliefs or desires.  Now, when the theory says that agents make certain sorts of transitions in their reasoning, we may need to rely on our ability to work with such contentful states themselves in order to trace out just where the particular initial beliefs and desires which the agent is characterized as having would lead her were she to make the sorts of generic transitions indicated.

To say that off-line processes have a crucial role in applying theory is not to say that simulation enters here.  It is to note that closely related cognitive strategies enter.  In simulation, off-line processes are employed with distinctive aspirations: to instantiate processes that in content and structure parallel those operative in the other.  The simulator aspires to do this by turning their own processes to work on those distinctive beliefs and desires possessed by the other, the content at transitions of the resulting process is supposed to be the same.  In contrast, in using off-line processes to apply theory, there need be no concern for strict parallels between the agent’s and the investigator’s processing.  The off-line processes are put to a different use—identification of applications of a theory to the case at hand.  In such cases, our own processes are employed off-line, but to trace out the implications of theory to the case at hand.

For concreteness, suppose that our theory indicates that folk with certain training will be insensitive to sample bias.  Our own processes may no longer run in such channels—due to different training.  However, we can trace out the contentful results of such free generalization by adding into the pretend-belief mix a pretend belief that the relevant sample is representative.  But, we should be under no allusion that the subject of our simulation has such a belief, or that it plays a role in freeing the subject to make the generalizations that we trace out.  It is not added in to our processing to parallel the agent’s processes, but to trace out where certain input would lead our subject, in keeping with our theory.  In any case, the place for something closely related to simulation in the application of theory should give us further reason to question the commonly imagined tension between these strategies.  Off-line cognition in the service of theory and theory-informed simulation are very close kin—the difference turning on how close a parallel in cognitive processing is aspired to.

In some cases, the off-line processes required to sustain the application of theory may themselves be plausibly viewed alternatively as small simulations.  Suppose that our theory includes some descendent to Festinger’s theory of cognitive dissonance.  It holds that, when a person has a set of beliefs and desires that are “dissonant,” that person will seek to revise his or her set in order to alleviate the dissonance.  Perhaps we are told that dissonance is related to inconsistency, but that a set of beliefs and desires can be dissonant without strictly being inconsistent.  One also experiences dissonance when recognizing or sensing that one’s set of beliefs and desires are inconsistent with the proposition that all is as it should be.   We get the general idea, perhaps aided by some examples.  Our own capacity for monitoring our own beliefs and desires for tension can then be put to work on beliefs and desires that we suppose someone to have.  This allows us to register when the antecedent conditions for dissonance resolution obtain in a given case—when, that is, the subject of our investigation has cognitive dissonance.  Then, our own abilities to explore sets of intentional states may be put to work, identifying various ways in which the dissonance might be exorcised by emending beliefs or modifying the values assigned to outcomes, and so on.  That is, our ability to work with such contentful states is put to work to identify some of the range of things that our subject might do in keeping with the theory.  Suppose that an agent has just been induced to part with a very significant sum in purchasing a new Volvo, under the influence of a persuasive sales person.  Suppose also that the agent is widely recognized as a tightwad, with the associated values and beliefs.   It is likely that we employ our cognitive capacities, off line, to identify this as a likely case of dissonance—we see that such an agent would readily have misgivings regarding whether all is now as it should be in his or her financial life.  We can project several ways in which the agent might then move to reassure him or herself—alleviating the dissonance.  We would not then be surprised to find our agent paying increased attention to safety data, or to reliability data in Consumer Reports.  These would allow the agent to reinforce his or her high evaluation of the new car.

In the application of the theory about cognitive dissonance and its effects, we have employed our own processes for identifying tension in our global belief sets, and for exploring ways of resolving tension.  Strictly speaking, for the application of this theory, there need be no suggestion that such processes parallel those operative in the agent—all that is required is that these processes identify for us likely cases of dissonance, and prominent routes to its resolution.  (Certainly there need be no suggestion that our new car buyer employs the notion of cognitive dissonance.)  However, in the present example, it is highly plausible that the off-line processes we employ are similar to those operative in our subject.  In such cases, we might conceive these processes interchangably as either simply the off-line application of theory or as stretches of theory-informed simulation.

Whatever the details, what the two cases illustrate is that, even when maximally relying on theory, the off-line use of our own capacity for working with contentful states may be needed in the course of applying theory.  If this is right, then theory with no help in its application from off-line processes will have very large blind spots due to the problems faced in its application.  It is appropriate to think of the maximal reliance on theory in terms of what would be needed to apply the theory effectively.  So, if our suggestions above are correct, the appropriate conception of the maximal reliance on theory in understanding, and minimal use of simulation, would allow for the use of off-line capacities for dealing with contentful states in tracing out the connections suggested by theory.  Here, these processes are used, not so much to parallel agents’ cognitive processing, as to apply our theory.  We think that it is fair to say that theory here can carry significant weight.

            Let us now note the blind spots to which the maximal reliance on theory will be subject.  Obviously, there are blind spots that result from the limitations of our psychological theory.  It is uncontestable that our psychological theory remains less than we would like.  Where it remains sketchy or silent, maximal reliance on theory without supplementation by simulation produces blind spots.

Some blind spots arising from the limitations of our current theories, will disappear as theory is improved.  So, some of the blind spots of maximal theory reliance can be expected to disappear.  However, we suspect that some blind spots will remain with us always.  For, as suggested above, a fully developed theory of relevance, is not to be had.  Where theory touches on cognitive processes that turn on the individual’s sensitivity to her global belief set, theory is and will remain sketchy.  This systematic and principled limitation on what is to be expected of theory gives rise to a systematic blind spot for the maximal reliance on theory.  As argued already, to epistemically move ahead with such cases, they must be handed over to simulation.

What would constitute a maximal reliance on simulation?  And what would be its blind spots?  Again, simulation is a matter of using one’s off-line cognitive processes to parallel the cognitive processes of the other, ascribing to the other parallel processes.  To initiate such a simulation, we must take certain of our cognitive processes off-line and provide them with select input.  The blind spots of maximal reliance on simulation are generally of two kinds: those that result from setting up the simulation in ill-informed ways, and those that result from the inflexibility of our cognitive processes.

Theory (or discursive information generally) may condition in epistemically important ways how one sets up a simulation.  Minimizing the role of theory can make for simulations that are less sensitive to the differences between persons.  Such lack of sensitivity makes for blind spots.  Theory may help one select which of one’s processes to employ when simulating certain others.  Theory, together with information about one’s biography and the biography of one’s subjects, may indicate that one has learned certain elaborations on rudimentary human cognitive tendencies that our subjects have not.  Accordingly, in simulating the other, one needs to employ the less tutored processes—if possible.  If one employs differently educated processes, errors will likely result.  Of course, one’s subjects may have been trained in ways that lead them to have processes diverging from both one’s trained processes and from one’s (presumably largely shared) rudimentary processes on which one’s earlier training worked.  Perhaps theory can inform the simulator how to adjust processes accordingly for purposes of the simulation.  Discursive information may also be important in determining the inputs to which subjects may be responding in a given context.  One would want such information to inform one’s pretend-belief and pretend-desire generator in setting up the simulation.  Without information about systematic differences in processes and inputs, simulation would become a more bumbling affair, one insensitive to differences between people.

At any point in time, a person’s cognitive processes are somewhat inflexible.  Yet, over time, with training cognitive processes can be shaped somewhat.  These simple points are the basis of an important limitation of simulation: my cognitive processes may not be such as can be made on demand to run on in ways parallel to those of various others. 

Our cognitive processes are undergoing a slow metamorphosis day in and day out.  Our teachers trained us; our colleagues continue to do the same.   We, and our experiences of success and failure contribute as well.  While slow, the changes can be cumulative.  After four years of college, one hopes that the students who have been our charges have acquired ways of reasoning that are significant refinements over those they possessed upon entering the academy.  Similar shaping of cognitive processes is pervasive.  Bird watchers do it; bee keepers do it; even educated and not so educated clergy do it.  And learning one way of reasoning is commonly unlearning others.  To learn a sensitivity to sample bias, one both learns to determine that a sample is (likely) representative before generalizing to the population and learns not to generalize to a population from samples without such a determination.  Since our rudimentary cognitive tendencies include tendencies to generalize from salient cases without attention to representativeness, this second process is unlearned while sensitivity to possible sample bias is learned.  In other contexts, or other communities, cognitive capacities may be trimmed and trained differently. 

            Suppose that training results in a certain set of transitions being natural to me.  For example, I may have learned to think of the probability of a given result in any one iteration of certain sorts of processes (like coin flips, or roles of die) as independent of earlier results, and as a function of enduring structural characteristics of the process (like the shape of the coin or die, and the distribution of its weight).  I am then unable to engage in the gambler’s fallacy.[10]  I may recognize how one committing that fallacy is reasoning, recognizing the contentful transitions in a person’s reasoning as an instance of a way of reasoning to which some folk are disposed.  (This is a significant applied theoretical understanding.)  But, if I take my cognitive processes off line, they don’t swing that way.  When I feed them the information that a coin has come up heads on the first five flips, and ask myself what to expect on the next flip, I get a question: “Was there something “funny” about that coin?”  I feel compelled to answer this question before answering the question posed.  When pretend-given that the coin was of an everyday sort pulled at random from someone’s pocket, I conclude that the coin is “fair” and that there is no basis for expecting heads or tails on the next flip.

            Our discussion of the DC (or dynamical cognition) framework gives us a useful way of conceiving the sort of cognitive inflexibility that concerns us here.  Trained in one way, my cognitive system has come to take on a certain topology. Trained in another way, a different person’s system has come to have a different topology.  The sequence of activation states will take a certain trajectory in my system, given what is “down hill” from a given input point.  The same input will take a different trajectory in the other, given that very different things may be “down hill” from that the point in activation space that represents that input in that system.  It may have taken a long time for each of us to acquire our respective cognitive topologies.  Changing that topology would likely be a slow process.  Changing my topology into the other’s, so that my processing upon input would really parallel the other’s, might be very difficult, if not impossible.  Certainly, it would not be a low-cost route to understanding the other.

            The picture that emerges from all this is of limited short-term flexibility in the cognitive processes that an individual can take off line together with somewhat greater plasticity in the long term, as training shapes the cognitive processes within each of our repertoires.  At any one time, there is some limited variation in the cognitive processes that I can call into play (either on line or off line).  Thus, the exact nature of my reasoning may vary as I find myself in various contexts at home with family, at professional presentations, in my office, in the gym, and so on.  This variability is a psychological fact.  Still, it is limited.  Another person, with different life experiences, will have come to have a somewhat different set of processes that can be called into play.  Presumably, there will be some overlap.  However, it also seems that, in each of our cases, there may well be processes that the other cannot call into play.  Successful simulation turns on the simulator’s cognitive repertoire including processes of the same sort as those that the subject of the simulation calls into play in the slice of life that is to be simulated.  Further, it turns on calling into play that process as opposed to others in the simulator’s limited repertoire.  As a result, maximal simulation faces two blind spots.  First, when the simulator’s repertoire includes processes of the sort that are in play in the simulated’s processing, but when the simulator calls into play different processes, simulation leads to misunderstanding.  Discursive information about individual biographies and about conditioned cognitive tendencies may play a role in setting up the simulation correctly.  Maximal simulation thus increases the risks of misunderstanding—needlessly expanding the scope of the blind spots to which simulation is subject.  So, the first set of blind spots faced by maximal simulation arise from its selective blindness (or near sightedness) with respect to differences in people and their cognitive processes.  Second, even allowing for the use of discursive information to condition simulation, it is generally subject to blind spots owing to the limited flexibility of cognitive processes.  When the processes in one’s repertoire do not include those that the subject of simulation employs, simulation is blind.[11]

            It is manifest that each of us can describe processes that we cannot readily employ.  Again, the simple case of the gambler’s fallacy provides a case in point.  It is easy enough to describe the fallacy.  It is also easy enough to recognize instances.  What is not easy, these days, is getting our own processes to freely run along those lines.  Of course, after a fashion, we can trace out the agent’s reasoning.  But, as seen earlier, what is done when I do this seems rather more like applying theory than like simulation proper.   Another example, one having more to do with desires, might be the compulsion that some reportedly experience with respect to certain forms of sadomasochism.

It is also plausible that one can simulate processes that one cannot describe.  For someone similar enough to ourselves, one can simulate being moved in certain ways by a movie or by a photograph, while having very little discursive understanding of the processes that are responsible.

We can begin to appreciate the epistemic importance of these observations while thinking in terms of the epistemic virtue of reliability.  To say that one has the capacity to simulate a range of human behavior and thought is to say that, as applied to that range of cases, simulation is reliable.  It is to say that simulation is reliable in environments where there is a preponderance of such cases.  To say that one can successfully apply theory, thereby accounting for a range of behavior and thought is to say that, so applying theory to the relevant cases is reliable—or that one’s application of that theory in an environment where it covers the preponderance of cases is reliable.  Regarding the epistemic benefits to be gotten from the joint application of simulation and theory, at least two conclusions now can be drawn.

First, insofar as both simulation and the application of theory can be reliable, each within its own somewhat varying range, applying them jointly, in a sensitive and coordinated manner, should be reliable in yet a greater range of cases than applying either of them in a maximally exclusive fashion.  While the application of theory by itself will be inadequate to deal with cases in which relevance within the cognizer’s global belief- and desire-sets is crucial, simulation provides a natural supplement.  Further, while theory will be limited in yet other ways at any one point in time, simulation allows us to reliably handle some of the cases where theory is inadequate.  On the other hand, simulation is intrinsically subject to limitations due to the inflexibility of any individual’s cognitive processes at a time.  Because theory will have applications to some of the resulting unsimulatable cases, it may extend the range of cases reliably handled.  In any case, theory, and discursive information generally, informs simulation in important ways, contributing to the range of cases in which simulation is itself reliable.  Thus, the competence resulting from the coordinated application of theory and simulation would seem to be reliable in a set of cases that includes (roughly) the union of the sorts of cases on which the two are individually reliable.

So, the coordinated application of both theory and simulation expands the range of sorts of cases with which we can reliably deal.  And, judging from the many true beliefs that we generate in our day-to-day dealings with folk, the partnership is reliable, and sufficiently so that there need be no objection on that score to enshrining it in our epistemic competence.

The gains in range of reliability attendant upon a cooperative use of theory and simulation have some implications for other epistemic virtues.  For example, cooperation seems a boon for systematicity.  First, simply the generation of a greater range of true beliefs about the behavior and mental life of individuals enrichens our set of beliefs.  We come to understand a more diverse set of individuals.  Second, because theory is here embraced as a useful component of our managing to understand others, it is natural to continue to develop theory so as to integrate this increasing diversity of cases into a yet more encompasing systematic treatment.

We get a complementary picture when we recast our observations in terms of  the reliability of the strategies when they are applied in differing environments.  We find that the cooperative application of theory and simulation has the virtue of being more robust than simulation or theory application alone.  For a striking comparison begin by reflecting on maximal simulation with a minimum of help from discursive information (we will want to note the effects of then refining simulation using discursive information and ultimately of complementing simulation with theory application).  Suppose that a person, call him Bob, has grown up in a highly homogeneous community, call it Russell, Kansas.   In our first possible world, Bob and most all his contemporaries remain in Russell throughout their lives, happy to ignore the wider world.  As a result, the great preponderance of cases to which Bob and other Russellians must turn their capacities for understanding are those involving people rather a lot like themselves in ethnic and family backgrounds, in formal and informal training, in beliefs and in life expectations.  Suppose that Bob and friends, being little interested in psychological theory, rely largely on simulation, and that their capacities for simulation constitute the bulk of their competence here.  Happily for them, simulation is fairly reliable in their environment.  For such a set of agents, inhabiting a world in which their cognitive environment is limited and largely populated by folk much like themselves, simulation is fairly reliable all by itself.  In fact, if the social environment is homogeneous enough, there seems little refinement for theory or discursive information to effect in setting up simulations.  Even maximal reliance on simulation with minimal use of discursive information–even untutored simulation (in which relatively little adjustment is made in processes to be taken off line and in beliefs and desires supposed in the simulation—would seem acceptably reliable.

Now, suppose that Bob lives in a somewhat different world.  In this world, as a young man Bob leaves Russell to join the military.  Here he is exposed to a more diverse set of people.  Perhaps, he returns home and uses his veteran’s benefits to pursue his education.  Perhaps he becomes interested in anthropology, or political science, or law.  In his subsequent years, he finds himself doing field work in strange places.  Or, he enters political life and attains high office in Washington.  In any case, in such worlds, Bob’s environment is wider, and the occasions for putting his capacity for understanding to work are more varied.  We may add that similar mobility characterizes many of Bob’s old Russellian cohorts—they are scattered by vicissitudes of the contemporary job market, family life, corporate transfers, and so on.  They also encounter a heterogeneous set of folk.  Here, their capacities for simulation will be less reliable, perhaps unacceptably so.  Until they bootstrap themselves into an understanding of the diversity they face, and until they employ such information both to inform their simulations, and to selectively curtail their reliance on simulation, they will often go wrong.  There are two ways that discursive information enhances the reliability of their practices, and thus two ways that maximal simulation with minimal reliance on discursive information is epistemically compromising.  First, simulation whose setup is minimally informed is less reliable than simulation need be.  Second, when an epistemic agent cannot “fall back” on the application of theory where cognitive inflexibility becomes limiting, the resulting overextended simulation is systematically misleading.  Until these agents realize that, at certain points, certain folk are difficult, if not impossible, to really simulate, their ready use of simulation will repeatedly lead them wrong.

Thus, as illustrated by the plight of the Russellians in differing worlds, the reliability of simulation will vary inversely with the heterogeneity in the cognitive processes to be encountered in a particular world or environment.  There are presumably many worlds in which there is significant diversity in folk’s social environment.  Accordingly, while the use of simulation alone may be reliable in some worlds, such as that in which Bob and his fellow Russellians stay at home, simulation by itself fails to be robust. 

We believe that this lack of robustness makes maximal simulation epistemically undesirable, even for the Russellians who happen to luck into a particular world in which it is reliable.  Given that the epistemic situation is intrinsically one of uncertainty regarding what the world is like, it is desirable to employ processes that would be reliable in a wide range of possible worlds.  To do otherwise is to court epistemic failure—it is to run an unnecessary risk of relying on unreliable processes.  Even if an agent happens to inhabit a particularly undemanding world, and thus lucks into employing processes with the virtue of reliability in that agent’s world, still, that agent’s cognitive processes nevertheless are unacceptably risky when lacking robustness.  Epistemically, it is not enough to be lucky; one should also be good—and this is to employ processes with the epistemic virtues of robustness and systematicity, as well as reliability.

On the other hand, the cooperative use of theory and simulation could be fairly robust.  In worlds where a folk face a cognitively homogeneous environment, simulation will be reliable as suggested above.  If anything, this reliability will be augmented by refinements brought about as the set up of simulation comes to be better informed with respect to what diversity there might be.  In worlds where they face a heterogeneous social environment, the development and application of theory promises both to refine simulation, where this is feasible, and to cover for the limitations of simulation where refinement is not to be had.  Theory, joined with simulation, seems the way to salvage reliability in the face of cognitive diversity (if it is to be salvaged).

Of course, our world seems to be characterized by a good deal of heterogeniety.  So, simulation alone fails to be either robust or reliable.  On the other hand, the joint application of theory and application may well be both robust and reliable, and it is at least congenial to systematicity.  We believe that such a joint application is reflected in the ways that we manage to understand others, and we conclude that this is central to our epistemic competence in understanding others.

V. On the Social Sciences

There is a long debate over whether the social sciences are best understood in terms of models of interpretation or models of explanation.  This debate has been misguided.  Its rotten fruit result from reliance on models of interpretation and explanation that are themselves deeply flawed.  As more adequate understandings emerge, the putative contrasting and competing approaches to the social sciences--the explanatist and interpretivist model--cease to conflict.  When explanation, for example, is no longer forced into the distorting mold of the covering-law model, we come to recognize that interpretivist social science serves up explanations.  Further, when we look at debates over the merits of competing interpretations, we find much turning on evaluations of the merits of the explanations that the competitors spawn.  Thus, successful explanation turns out to be a desideratum for correctness of interpretation.  At the same time, explanations turn on interpretations--from which their understanding of initial conditions, or their pretend-beliefs, are taken.  Accordingly, our confidence that any given explanation is correct should itself partially turn on the successes that can be claimed for the interpretive scheme on which that explanation draws.  Accordingly, an interpretivist model of the social sciences must also be an explanationist model, and an explanation-centered model should also be an interpretivist model (see Henderson 1993, and Risjord forthcoming).

There has been some tendency to see the clash between simulation theorists and theory theorists against the background of the older debate, with theory theorists viewed as partisans of the explanation-centered understanding and simulation theorists viewed as championing yet another interpretivist model.  This is doubly mistaken.  It compounds the mistakes of the older debate by somehow forgetting that the simulation theorist seek to model our intentional explanations of folk.  The mistake is understandable, as the talk of applying theory suggests to some the sort of subsumption under covering-laws that once provided the philosopher’s model of explanation--but that time should be past.  Both theory theory and simulation theory attempt to model our explanations of others in intentional terms.  If we are correct here, both manage to model some of our successful explanatory practice.  Both, are partial models of our epistemic competence for intentional explanations.  The full model requires a hybrid model--one providing for the coordinated application of these strategies.  To emphasize, Henderson (1995) shows that these models should not be taken as competing accounts of explanations--describing two very unlike sorts of putative explanation--but as characterizing different epistemic routes to a single sort of explanation, one in which the causal antecedents of the action to be explained are identified (in response to a why-question).  Here, we have elaborated these themes by showing how each strategy constitutes one component of our epistemic competence for generating such explanations.  We have argued that the coordinated use of the two strategies has the full range of epistemic virtues that we seek in our cognitive practices: it is acceptably reliable, is congeniel to systematicity, and is robust.

Even so, we feel compelled to insist that even the account of our epistemic competence advocated here falls short of a full picture of the human sciences (a point that should be obvious).  It is an account of our ability to generate explanations, or possible explanations, drawing on background theory and on background information about people.  Sometimes we apply theory.  Sometimes we generate (in an informed way) pretend beliefs and desires.  Sometimes we do some of both.  In any case, when what we are then led to matches, so far as we can tell, what is observed, then we have a possible explanatory success. When we are led in ways do not match what seems to be observed, we know that either our theory was wrong (if we were applying significant theory), or that our information about our subject's beliefs and desires was wrong, or that we did not manage to call into play cognitive capacities quite like those that our subject employed.  (We noticed this range of pitfalls earlier.)   In such cases, we must make informed choices of where to tinker.  We may tinker with our theoretical understandings of human cognitive capacities, the forms they can take, and their plasticity with training and with situations.  We may tinker with our putative information regarding the folk in question. In emending what we take our subjects to believe or desire, we commonly are also tinkering with our interpretive scheme for them.  A full epistemic understanding of our practice in the social sciences would deal with our capacities for refining and developing theory (see for example, Henderson 1993, chpt. 7), and with our capacities for building and refining interpretive schemes in the face of explanatory successes and failures (Henderson, 1993, chpts. 1-3; Henderson, 1994b; Risjord, forthcoming).  Ultimately, then, a full understanding of the epistemic competence that underlies our social sciences will need to embed the capacity for coordinated use of  theory and simulation as but one component, albeit an important one.

 

References

Berg,  G.  1992. “A Connectionist Parser with Recursive Sentence Structure and Lexical Disambiguation.”  In AAAI-92: Proceedings of the Tenth National Conference on Artificial Intelligence.  AAAI Press/M.I.T. Press.

Evans-Pritchard, Edward. 1937. Witchcraft, Magic and Oracles Among the Azande. Clarendon Press.

Fodor, Jerry. 1983. The Modularity of Mind: An Essay in Faculty Psychology. M.I.T. Press.

Grandy, Richard. .1973. “Reference, Meaning, and Belief.” Journal of Philosophy 70: 439-452.

Goldman, Alvin. 1989. “Interpretation Psychologized.” Mind and Language 4:161-185.

Heal, Jane. 1996. “Simulation, Theory, and Content.”  In P. Carruthers and P. Smith, eds.  Theories of Theories of Mind.  Cambridge University Press, pp. 75-89.

Henderson, David. 1993. Interpretation and Explanation in the Human Sciences.  State University of New York Press.

________. 1994b. "Conceptual Schemes After Davidson." In G. Preyer, et al., eds. Language, Mind, and Epistemology. Kluwer.

________,  1994a. “Epistemic Competence.” Philosophical Papers 23: 139-67.

________. 1995. “Simulation Theory vs. Theory Theory.” Southern Journal of Philosophy 34: 65-93.

Henderson, David and Horgan, Terence.  Forthcoming. "Iceberg Epistemology."  Philosophy and Phenomenological Research.

Henderson, David and Horgan, Terence.  MS. “Practicing Safe Epistemology.”

Horgan, Terence.  1997. “Connectionism and the Philosophical Foundations of Cognitive Science.” Metaphilosophy 28: 1-30.

Horgan Terence. and Tienson, John. 1990.  “Soft Laws.” Midwest Studies in Philosophy 15: 256-79.

________. 1994. “A Nonclassical Framework for Cognitive Science.” Synthese 101: 305-45.

________. 1995. “Connectionism and the Commitments of Folk Psychology.” Philosophical Perspectives 9: 127-52.

________. 1996. Connectionism and the Philosophy of Psychology.  M.I.T. Press.

Humphreys, Paul. 1989. The Chances of Explanation.  Princeton Univ. Press.

Kitcher, Philip. 1992. “The Naturalist Returns.” Philosophical Review, 101: 53-114.

Pollack,  J. 1990. “Recursive Distributed Representations.” Artificial Intelligence 46: 77-105.

Risjord, Mark. forthcoming: Woodcutters and Witchcraft. State University of New York Press.

Smolensky, Paul. 1990.  “Tensor Product Variable Binding and the Representation of Symbolic Structures in Connectionist Systems.” Artificial Intelligence 46: 159-216.

________.  1995.  “Connectionist Structure and Explanation in an Integrated Connectionist/Symbolic Cognitive Architecture.”  In C. MacDonald and G. MacDonald, eds. The Philosophy of Psychology: Debates on Psychological Explanation, Volume Two, 357-94.  Blackwell.