There is relatively little work in cognitive science that investigates the nature of heuristics and the mechanics of their operations. This paper outlines a theory of how heuristics work in cognition. A cognitive architecture is described which facilitates satisficing procedures which utilize little cognitive resources, i.e., heuristics. An interesting feature of this account is that heuristics are not assumed to operate with representational content, but with higher-order information (or knowledge) implicit in informationally rich representational systems and relations among them.
Heuristics are commonly understood as economical shortcut reasoning procedures that may not lead to optimal or correct results, but will generally produce outcomes that are in some sense satisfactory or “good enough”. Since the seminal work of Kahneman and Tversky in the 1970s, there has been an explosion of research that tries to uncover various heuristics that humans typically rely on in problem-solving or decisionmaking. Yet there has been comparatively little research on the general nature of heuristics and how they operate in cognition.
An exception is the recent incarnation of heuristics
research in psychology, exemplified by Gigerenzer’s “fast and
frugal heuristics” research program. Gigerenzer and his
colleagues make various attempts to explain how heuristics work
by proposing computational models and exposing the
information that a given heuristic is sensitive to. However, it is
not clear whether human cognition actually implements the
proposed computational models
Even less philosophical attention has been paid to the nature of heuristics and heuristic reasoning. Yet philosophy as a discipline possesses valuable resources to advance investigations into these important matters, such as accounts of cognitive architecture, theories of representation, theories of concepts, and theories of reason and cognition.
In this paper I outline a general theory of how
heuristics work in human cognition which enjoys inspiration from
philosophical reflections on cognitive archtiecture. While
heuristics research in the psychological literature focuses on
determining what specific information or knowledge is used
by given heuristics, I will here reorient the problem by
discussing architectural properties of the mind that facilitate
heuristics generally. More specifically, I will investigate
cognitive architectural features that can constrain heuristic
processes in such a way that ensures that they satisfice, operate
systematically, and require little cognitive resources for their
recruitment and execution
Let us begin with some informative remarks from Dennett to guide our discussion. In his paper, “Cognitive Wheels”, he writes:
Even if you have excellent knowledge (and not mere
belief) about the changing world, how can this knowledge
be represented so that it can be efficaciously brought to
bear? . . . A walking encyclopedia will walk over a cliff,
for all its knowledge of cliffs and the effects of gravity,
unless it is designed in such a fashion that it can find the
right bits of knowledge at the right times, so it can plan
its engagements with the real world.
In a vein similar to Dennett’s remarks, some theorists
believe that a modular architecture enables fast and frugal
heuristics
Informational encapsulation is supposed to confine the
amount of information that can be surveyed to a highly
restricted proprietary database, which significantly reduces the
computational burden associated with information search.
Nevertheless, there are at least two reasons to doubt that
heuristics owe their ability to be fast and frugal to
informational encapsulation. First, heuristics are pervasive in central
cognition, and yet central systems don’t exhibit modular
encapsulation: a widely recognized feature of central systems is
that they allow for the free exchange of information,2 which is
antithetical to encapsulation
On the other hand, perhaps it is domain-specificity that
enables heuristics to operate fast and frugally. Indeed, the
potential benefits of domain-specificity and its role in cognition
are especially apparent. As Samuels explains, “if a
mechanism is sufficiently domain specific, then it becomes possible
to utilize a potent strategy for reducing computational load,
namely, to build into the mechanism substantial amounts of
information about the domain in which it operates”
Nevertheless, if we are to take the lessons from Dennett seriously, then we must add to this that the information built into a system must be encoded in a highly organized fashion. A mechanism can take advantage of generous amounts of domain-specific information encoded in a system, and thereby enable quick and efficient cognition, but if the information is structured or organized in specific ways, it would enable quicker and more efficient cognition. Contrariwise, if such information is not structured or organized in specific ways, it would retard the speed and efficiency of search and processing. Characteristic human performance on many cognitive tasks suggests that the structures heuristics exploit are distinctly organized to produce systematicity and robustness in reasoning and inference.
Moreover, it would seem that a system must not only have a
specific internal structure, but also bear specific intra-system
relations, which would further facilitate quick and efficient
reasoning. Reasoning within one domain often bears in a
1These two properties were, of course, those (among others) that
2This is a feature acknowledged by those who deny central
systems modularity
3Of course, there may be other reasons why an encapsulated system might deploy heuristics other than information search (e.g., pattern matching), but the claim that I’m trying to make here is that informational encapsulation and heuristics can be divorced in such a way that warrants resisting the idea that heuristics satisfice and require little cognitive resources in virtue of informational encapsulation. number of ways on other domains. Without such connections between bodies of knowledge, we would not be able to make the rich heuristic inferences we characteristically do.
But for all this, there is no need to refer to these organized systems as “domain-specific”. It seems rather that what is really doing the work in facilitating fast and frugal cognition is not domain-specificity per se, but the manner in which the specific kinds of information is encoded and organized.4 For, similar to what was recently observed, it is possible that one can have an unorganized system of lots of domain-specific information, but it will be doubtful that a cognitive mechanism would exhibit the same speed and efficiency operating with this unorganized body of knowledge as it would operating with a highly structured system. Moreover, it is possible that one can have a domain-specific body of knowledge that is quite impoverished; and although a cognitive mechanism would likely be able to operate quickly and efficiently with such a system, owing primarily to the little information that ever gets considered, the inferences made would not be as robust as those made by operations over richer systems.
In summary, a plausible feature of cognitive architecture that would enable the fast and frugal operations of heuristics is the possession of numerous highly-organized epistemic systems that allow easy access to certain information in certain circumstances. Since neither informational encapsulation nor domain-specificity really play a role here, it would be improper to call these systems modules. But what, then, could these systems be? It turns out that this philosophicallyinspired analysis converges on some important psychological models of cognition.
There is a long tradition in psychology which describes
human representational systems in terms of networks. These
models have roots in semantic processing theories,
according to which information about words and their meanings
are stored and organized in complex webs of nodes. Nodes
are understood to be specific representations—variously
conceived as concepts, features, or lexical items—with
connections that reflect associative and/or semantic relations. (See
Such network models have been used as paradigms to
investigate conceptual cognition, memory recall and
organization, semantic priming, and the properties of lexical systems
Nevertheless, the usual characterizations of network nodes
are insufficient for an account of cognition. As recently
stated, nodes are usually understood variously as concepts,
features, or lexical items; and qua nodes, they are
generally conceived to be unstructured units. Now, lexical items
really seem to be unstructured units, and as such their
natures and roles in cognition may very well be captured
by network models. However, the embodiment view of
concepts or semantics maintains that semantic
representations (at least partly) consist of sensory-motor
representations
As I am advancing a theory of how heuristics work, these
considerations urge me to recast network models for
spreading activation in such a way that nodes are conceived to be
structured units consisting of numerous multimodal
representations
Conceiving network nodes in this way seems to be precisely what is needed for the philosophically-inspired analysis of cognitive architecture above. The main lesson there was that highly organized epistemic structures are needed to facilitate fast and frugal cognition. In light of what has just been discussed, the internal structure of nodes together with the structure of the embedding network appear to provide the right sort of structures to facilitate access to salient representations (to be used to make inferences and judgments) by means of spreading activation.
Moreover, it would be worthwhile to maintain the
features of many network models mentioned above, namely that
the connections between nodes are directed and weighted
Consider an example. Activating representations of cats might almost always activate representations of furriness for an individual, but activating representations of furriness might activate representations of cats only in certain contexts (the directed connection is heavily weighted from cats to furriness, and the directed connection is lighter the other way). Furthermore, if this individual is very familiar with cats, representations of furriness might be more readily activated relative to representations of purr, even though both furriness and purr bear directed connections from cats. This will entail 6Henceforth, when I refer to representations I mean to refer to multimodal representations, even though I omit explicit mention of their multimodal character. that, for this individual across contexts involving cat representations, cats’ furriness will in general be more salient than their purring. In short, these features add to the organized structure of one’s representational system that, as we saw, is so important to facilitate access to the right representations at the right times.7
When faced with a cognitive task, the relations among nodes
in a network embody a rich source of implicit (putative)
knowledge concerning the active representations, but which
is not contained within the representations themselves. Such
implicit information includes which representations are
associated with others, structural or hierarchical relation
information among representations
To illustrate, let us briefly consider Gigerenzer’s Take the
Best heuristic
Gigerenzer’s explication of Take the Best assumes that an individual has a subjectively ranked order of beliefs about cues that may discriminate between objects, and which are considered sequentially when making predictions about the object along certain dimensions. The highest ranked cue which discriminates is termed “the best”, and the object that has a positive value with respect to this best cue will be predicted to possess some criterion. For example, suppose you had to choose which of a pair of cities is larger. You might take recognition as a cue that would discriminate and thereby inform your choice. However, if both cities are recognized, you move down your ordered list of cues to the next one that discriminates. Suppose that having a university is the next cue 7This sort of account of cognitive architecture is notably different from what is assumed by classical architectures such as ACT-R and Soar. The latter sorts of architectures assume that cognition consists of operations over propositions, and further that rules of inference (e.g., productions) are explicitly represented propositionally. On the present account, there is no assumption that cognition is propositional, even though the account is compatible with this assumption. Further, there is no assumption that heuristics are rule-like, even though (again) the account is compatible with rule-governed operations. Thus, the present account has certain advantages over classical architectures insofar as it is more general, and can accommodate non-classical architectures as well as multimodal representations. In addition, unlike what is needed for ACT-R or Soar, the present account does not assume that problems and problem-spaces are or must be well-defined.
8I say “primarily” since some representational content may
figure into some heuristic processes. My point, nonetheless, is that
representational content is not what heuristics generally operate with.
that discriminates on the criterion of city size, and you know
that one city has a university while the other does not. Having
a university would be “the best” cue on this occasion—you
would thus use Take the Best and infer that the city with the
university is the larger of the two
To explain how this heuristic works in terms of the account outlined in the present paper, we begin by observing that a set of representations will be activated by the cognitive task. Network nodes that might be active in this task can include clusters of representations for city, infrastructure, roads, traffic, buildings, university, among others; the sorts of nodes and representations that get activated in a given task is open-ended depending on the representational and conceptual wherewithal of the individual. Now, it is important to understand that certain beliefs are implied about the cue upon which the choice in question is made. At the very least, we can infer that the said cue was believed to be “the best” on which to make the choice in question, and believing that cue to be “the best” implies certain things about the representational content of the cue (as possessed by the chooser), as well as certain things about how the cue fits within the chooser’s representational system. For instance, if one believes that having a university is a good predictor of relative city size, one’s university representations must be rich enough to support such a belief, or at least to support the belief that having a university is a better predictor of relative city size than some other cue (e.g., having a major highway). All of this knowledge is conceived as constituting a particular node as a cluster of university representations (as indicated above). Implicit information built into one’s university node might include, in addition to various representations and conceptualizations that are required to understand what a university is, information about a university’s function in society; how it houses a number of faculty, staff, and students; perhaps the relative social classes of members of a community typical of such faculty, staff, and students; maybe the ways in which having a university relates to the economy of a city; and probably much more. Indeed, it is such information that generally makes one believe that having a university is a good predictor of relative city size. In other words, this implicit knowledge will help guide one’s reasoning to infer that the city with the university is a large city, or at least larger than the other. (We might recall here the lesson from Dennett discussed above.)
The diagram in Figure 1 is a rough characterization of a possible network that would be active in the Take the Best task we have been discussing. The structure exhibited suggests to the chooser that having a university, being a capital city, having an airport, and having a major highway run through the city are implicitly believed to indicate relative city size. Take the Best can thus be viewed as operating with the suggested structure to deliver an inference that the city in question is the larger of the two under consideration. In general, heuristics are conceived to exploit active structures made available through spreading activation (i.e., heuristics don’t operate over non-active connections). Note in the Figure the multiple, heavily-weighted connections pointing to the large city node—the Take the Best heuristic would exploit this information made available through the spreading activation process.
On this view, we might thus understand one’s belief that
having a university is a good predictor of relative city size to
be implicit in metainformational structures that exist among
representations comprising the network. In this way, beliefs
about discriminating cues need not be explicitly represented.
This is contrary to Gigerenzer’s view that the belief that a
cue discriminates is explicitly represented and searched. Also
contrary to Gigerenzer’s view is that heuristics are here
understood to operate by generally ignoring representational
content. Instead, heuristics utilize cognitive structures (viz.
connections that exist among nodes and representations) that
implicitly embody higher-order information (or knowledge)
about active representations and their content9
In addition, on the present account, it will be very
common for one to activate lots of conceptual content. This is
because the cognitive tasks we typically face have a high
cognitive load, and as such, navigating these tasks invoke large
amounts of representations and informationally rich
connections among them. This is antithetical to Gigerenzer’s
assumption that heuristic decision-making is often based on one
reason
Importantly, despite the exploitation of generous amounts
of information, none of this implies that heuristics are not
frugal. Frugality is a key feature of heuristics, and moreover,
heuristics are supposed to satisfice
The theory outlined in this paper was advanced to explain how heuristics work in cognition. It is my hope that it will also serve to motivate researchers to take seriously the need to better understand what cognitive heuristics are and how they operate. Since heuristics are supposed to play a large role in our cognitive lives, our picture of human cognition will remain incomplete without such an understanding.
I acknowledge that more work needs to be done to
corroborate the foregoing understanding of how heuristics work. I
(all too briefly) illustrated how the Take the Best heuristic can
be understood in terms of the theory outlined in this paper.
Other heuristics in the literature can be similarly understood,
but it is not possible to show this here. Further research must
be done to explore more fully the nature of our
representational structures and systems, the nature of the relations
between representations and representational content, as well as
10
I would like to thank Richard Samuels, Chris Viger, and the anonymous referees of EAPCogSci2015 for helpful comments and suggestions.