CHAPTER 1

A Short Primer on Situated Cognition

Philip Robbins and Murat Aydede

In recent years there has been a lot of buzz about a new trend in cognitive science. The trend is associated with terms like embodi-ment, enactivism, distributed cognition, and the extended mind. The ideas expressed using these terms are a diverse and sundry lot, but three of them stand out as especially central. First, cognition depends not just on the brain but also on the body [the embodi-ment thesis]. Second, cognitive activity rou-tinely exploits structure in the natural and social environment (the embedding thesis). Third, the boundaries of cognition extend beyond the boundaries of individual organ-isms (the extension thesis). Each of these theses contributes to a picture of mental activity as dependent on the situation or context in which it occurs, whether that sit-uation or context is relatively local (as in the case of embodiment) or relatively global (as in the case of embedding and extension). It is this picture of the mind that lies at the heart of research on situated cognition. According to our usage, then, situated cognition is the genus, and embodied, enactive, embedded, and distributed cognition and their ilk are species. This usage is not standard, though

it seems to us as good as any (for compet-ing proposals, see Anderson, 2003; Clancey, 1997; Wilson, 2002).

In this brief introductory chapter, we present a bird's-eye view of the concep-tual landscape of situated cognition as seen from each of the three angles noted previ-ously: embodiment, embedding, and exten-sion. Our aim is to orient the reader, if only in a rough and preliminary way, to the sprawling territory of this handbook.

1. The Embodied Mind

Interest in embodiment - in "how the body shapes the mind," as the title of Gallagher (2005) neatly puts it - has multiple sources. Chief among them is a concern about the basis of mental representation. From a foun-dational perspective, the concept of em-bodiment matters because it offers help with the notorious "symbol-grounding prob-lem," that is, the problem of explaining how representations acquire meaning (Anderson, 2003; Harnad, 1990; Niedenthal, Barsalou, Winkielman, Krauth-Gruber, & Ric, 2005).

3

4 PHILIP KOBB1NS AND MURAT AYDEDE

This is a pressing problem for cognitive sci-ence. Theories of cognition are awash in representations, and the explanatory value of those representations depends on their meaningfulness, in real-world terms, for the agents that deploy them. A natural way to underwrite that meaningfulness is by grounding representations in an agent's capacities for sensing the world and acting in it:

Grounding the symbol for 'chair', for instance, involves both the reliable detec-tion of chairs, and also the appropriate reactions to them The agent must know what sitting is and be able to systemati-cally relate that knowledge to the perceived scene, and thereby see what things (even if non-standardfy) afford sitting. In the nor-mal course of things, such knowledge is gained by mastering the skill of sitting (not to mention the related skills of walking, standing up, and moving between sitting and standing), including refining one's per-ceptual judgments as to what objects invite or allow these behaviors; grounding 'chair', that is to say, involves a very specific set of physical skills and experiences. (Anderson, 2003, pp. J02-03J

This approach to the symbol-grounding problem makes it natural for us to attend to the role of the body in cognition. After all, our sensory and motor capacities depend on more than just the workings of the brain and spinal cord; they also depend on the work-ings of other parts of the body, such as the sensory organs, the musculoskeletal system, and relevant parts of the peripheral nervous system (e.g., sensory and motor nerves). Without the cooperation of the body, there can be no sensory inputs from the environ-ment and no motor outputs from the agent -hence, no sensing or acting. And without sensing and acting to ground it, thought is empty.

This focus on the sensorimotor basis of cognition puts pressure on a traditional con-ception of cognitive architecture. According to what Hurley (1998) calls the "sandwich model," processing in the low-level periph-eral systems responsible for sensing and act-ing is strictly segregated from processing in

the high-level central systems responsible for thinking, and central processing oper-ates over amodal representations. On the embodied view, the classical picture of the mind is fundamentally flawed. In particu-lar, that view is belied by two important facts about the architecture of cognition: first, that modality-specific representations, not amodal representations, are the stuff out of which thoughts are made; second, that perception, thought, and action are co-constituted, that is, not just causally but also constitutively interdependent (more on this distinction follows).

Supposing, however, that the sandwich model is retired and replaced by a model in which cognition is sensorimotor to the core, it does not follow that cognition is embod-ied in the sense of requiring a body for its realization. For it could be that the sensori-motor basis of cognition resides solely at the central neural level, in sensory and motor areas of the brain. To see why, consider that sensorimotor skills can be exercised either on-line or off-line (Wilson, 2002). On-line sensorimotor processing occurs when we actively engage with the current task envi-ronment, taking in sensory input and pro-ducing motor output. Off-line processing occurs when we disengage from the envi-ronment to plan, reminisce, speculate, day-dream, or otherwise think beyond the con-fines of the here and now. The distinction is important, because only in the on-line case is it plausible that sensorimotor capacities are body dependent. For off-line function-ing, presumably all one needs is a working brain.

Accordingly, we should distinguish two ways in which cognition can be embodied: on-line and off-line (Niedenthal et al., 2005; Wilson, 2002). The idea of on-line embodi-ment refers to the dependence of cogni-tion - that is, not just perceiving and acting but also thinking - on dynamic interactions between the sensorimotor brain and rele-vant parts of the body: sense organs, limbs, sensory and motor nerves, and the like. This is embodiment in a strict and literal sense, as it implicates the body directly. Off-line embodiment refers to the dependence

A SHORT PRIMER O N SITUATED COGNITION 5

of cognitive function on sensorimotor areas of the brain even in the absence of sen-sory input and motor output. This type of embodiment implicates the body only indi-rectly, by way of brain areas that process body-specific information (e.g., sensory and motor representations).

To illustrate this distinction, let us con-sider a couple of examples of embodiment effects in social psychology (Niedenthal et al., 2005). First, it appears that bodily postures and motor behavior influence eval-uative attitudes toward novel objects. In one study, monolingual English speakers were asked to rate the attractiveness of Chinese ideographs after viewing the latter while performing different attitude-relevant motor behaviors (Cacioppo, Priester, & Bernston, 1995). Subjects rated those ideo-graphs they saw while performing a posi-tively valenced action (pushing upward on a table from below) more positively than ideographs they saw either while performing a negatively valenced action (pushing down-ward on the tabletop) or while performing no action at all. This looks to be an effect of on-line embodiment, as it suggests that actual motor behaviors, not just activity in motor areas of the brain, can influence atti-tude formation.

Contrast this case with another study of attitude processing. Subjects were presented with positively and negatively valenced words, such as love and hate, and asked to indicate when a word appeared either by pulling a lever toward themselves or by pushing it away (Chen & Bargh, 1999). In each trial, the subject's reaction time was recorded. As predicted, subjects responded more quickly when the valence of word and response behavior matched, pulling the lever more quickly in response to posi-tive words and pushing the lever away more quickly in response to negative words. Embodiment theorists cite this finding as evidence that just thinking about some-thing - that is, thinking about something in the absence of the thing itself - involves activity in motor areas of the brain. In par-ticular, thinking about something positive, like love, involves positive motor imagery

(approach), and thinking about something negative, like hate, involves negative motor imagery (avoidance). This result exempli-fies off-line embodiment, insofar as it sug-gests that ostensibly extramotor capacities like lexical comprehension depend to some extent on motor brain function - a mainstay of embodied approaches to concepts and categorization (Glenberg & Kaschak, 2002; Lakoff & Johnson, 1999).

The distinction between on-line and off-line embodiment effects makes clear that not all forms of embodiment involve bodily dependence in a strict and literal sense. Indeed, most current research on embodi-ment focuses on the idea that cognition depends on the sensorimotor brain, with or without direct bodily involvement. (In that sense, embodied cognition is something of a misnomer, at least as far as the bulk of research that falls under this heading is concerned.) Relatively few researchers in the area highlight the bodily component of embodied cognition. A notable exception is Gallagher's (2005) account of the distinc-tion between body image and body schema. In Gallagher's account, a body image is a "system of perceptions, attitudes, and beliefs pertaining to one's own body" (p. 24), a complex representational capacity that is realized by structures in the brain. A body schema, on the other hand, involves "motor capacities, abilities, and habits that both enable and constrain movement and the maintenance of posture" (p. 24), much of which is neither representational in charac-ter nor reducible to brain function. A body schema, unlike a body image, is "a dynamic, operative performance of the body, rather than a consciousness, image, or conceptual model of it" (p. 32). As such, only the body schema resides in the body proper; the body image is wholly a product of the brain. But if Gallagher is right, both body image and body schema have a shaping influ-ence on cognitive performance in a variety of domains, from object perception to lan-guage to social cognition.

So far, in speaking of the dependence of cognition on the sensorimotor brain and body, we have been speaking of the idea that

6 PHILIP ROBBINS A N D MURAT AYDEDE

certain cognitive capacities depend on the structure of either the sensorimotor brain or the body, or both, for their physical real-ization. But dependence of this strong con-stitutive sort is a metaphysically demand-ing relation. It should not be confused with causal dependence, a weaker relation that is easier to satisfy (Adams & Aizawa, 2008; Block, 2005). Correlatively, we can distin-guish between two grades of bodily involve-ment in mental affairs: one that requires the constitutive dependence of cognition on the sensorimotor brain and body, and one that requires only causal dependence. This distinction crosscuts the one mooted ear-lier, between on-line and off-line embodi-ment. Although the causal/constitutive dis-tinction is less entrenched than the on-line/ off-line distinction, especially outside of phi-losophy circles, it seems no less funda-mental to an adequate understanding of the concept of embodiment. To see why, note that the studies described previously do not show that cognition constitutively depends on either the motor brain or the body. The most these studies show is some sort of causal dependence, in one or both directions. But causal dependencies are rel-atively cheap, metaphysically speaking. For this reason, among others, it may turn out that the import of embodiment for foun-dational debates in cognitive science is less revolutionary than is sometimes advertised (Adams & Aizawa, 2008).

2. The Embedded Mind

It seems natural to think of cognition as an interaction effect: the result, at least in part, of causal processes that span the boundary separating the individual organism from the natural, social, and cultural environment. To understand how cognitive work gets done, then, it is not enough to look at what goes on within individual organisms; we need to consider also the complex transactions between embodied minds and the embed-ding world. One type of such a transaction is the use of strategies for off-loading cognitive work onto the environment, a useful way to

boost efficiency and extend one's epistemic reach.

One of the best articulations of the idea of cognitive off-loading involves the concept of epistemic action (Kirsh & Maglio, 1994). An epistemic action is an action designed to advance the problem solver's cause by revealing information about the task that is difficult to compute mentally. The best-known example of epistemic action involves the computer game Tetris, the goal of which is to orient falling blocks (called "zoids") so they form a maximally compact layer at the bottom of the screen. As the rate of fall accelerates, the player has less and less time to decide how to orient each block before it reaches the bottom. To cope better with this constraint, skilled players use actual physical movements on the keyboard to manipulate the blocks on the screen - a more efficient strategy than the "in-the-head" alternative of mentally rotating the blocks prior to ori-enting them on the screen with keystrokes. A roughly analogous strategy of cognitive off-loading facilitates more mundane tasks like grocery packing (Kirsh, 1995). The prob-lem here is to arrange things so that heavy items go on the bottom, fragile items on top, and intermediate items in between. As the groceries continue to move along the con-veyor belt, decisions about which items go where need to be made swiftly, to avoid pile-ups and clutter. As items come off the con-veyor belt and enter the work space, skilled grocery packers often rapidly sort them by category (heavy, fragile, intermediate) into distinct spatial zones prior to placing each item in a bag. This procedure significantly decreases load on working memory relative to the alternative of mentally calculating the optimal placement of each item as it enters the work space, without the benefit of exter-nal spatial cues.

Both of these examples of epistemic action point to the importance of minimiz-ing load on internal memory, on working memory in particular. This echoes the twin themes of Brooks's (1991) "world as its own model" (p. 140) and O'Regan's (1992) "world as an outside memory" (p. 461). The com-mon idea here is that, instead of building

A SHORT PRIMER O N SITUATED COGNITION 7

up detailed internal models of the world that require continuous and costly updat-ing, it pays to look up relevant informa-tion from the world on an as-needed basis. In other words, "rather than attempt to mentally store and manipulate all the rele-vant details about a situation, we physically store and manipulate those details out in the world, in the very situation itself' (Wilson, 2002, p. 629). The suggestion that intelligent agents do best when they travel informa-tionally light, keeping internal representa-tion and processing to a minimum, informs a wide spectrum of research on cognition in the situated tradition (Clark, 1997). Vision science affords a nice example of this trend in the form of research on change blind-ness. This is a phenomenon in which viewers fail to register dramatic changes in a visual scene - a phenomenon that some interpret as evidence that the visual system creates only sparse models of the world, giving rise to representational blind spots (O'Regan, 1992).

The embedding thesis, then, goes hand in hand with what Clark (1989) calls the "007 principle."

In general, evolved creatures will neither store nor process information in costly ways when they can use the structure of the envi-ronment and their operations upon it as a convenient stand-in for the information-processing operations concerned. That is, know only as much as you need to know to get the job done. (p. 64J

Embedding, in turn, goes hand in hand with embodiment, as off-loading cognitive work depends heavily on sensorimotor capacities such as visual lookup, pattern recognition, and object manipulation. Epistemic actions, for instance, typically require embodiment in a strict and literal sense, as they involve real-time dynamic interaction with the local physical environment.

The theoretical and methodological import of embedding, however, is much wider. It points t o the importance, in gen-eral, of studying cognition "in the wild," with careful attention to the complex inter-play of processes spanning mind, body,

and world (Hutchins, 1995). The scope of this ecological perspective on the mind is very broad indeed. Having expanded far beyond Gibson's (1979) work on vision, it informs research programs in virtually every area of psychology, from spatial naviga-tion to language acquisition to social cog-nition. It is nicely illustrated by theories of social rationality, which try to explain human judgment and decision making in terms of the structure of the social envi-ronment (Gigerenzer, 2000). Somewhat fur-ther afield, the ecological view has begun to show up with increasing frequency in the literature on phenomenal consciousness, that is, consciousness in the "what-it's-like" sense popularized by Nagel (1974). It is implicit, for example, in the enactivist idea that the felt quality of visual awareness is a by-product of ongoing agent-environment interaction (No, 2004). It also informs con-structivist conceptions of consciousness, such as the idea that an individual's con-scious mental life tends to mirror that of socially salient others (Robbins, 2008). Both of these suggestions about the nature of phe-nomenal consciousness - arguably the last bastion of Cartesian internalism - reflect a newly invigorated ecological perspective on the mind.

3. The Extended Mind

Assigning an important explanatory role to brain-body and agent-environment interac-tions does not constitute a sharp break from classical cognitive science. Both the embodi-ment thesis and the embedding thesis can be seen as relatively modest proposals, given that they can be accommodated by rela-tively minor adjustments to the classical pic-ture, such as the acknowledgment that "not all representations are enduring, not all are symbolic, not all are amodal, and not all are independent of the sensory and effector sys-tems of the agent" (Markman & Dietrich, 2000, p. 474; see also Vera & Simon, 1993). The same cannot be so easily said, however, of the claim that cognition is extended -the claim that the boundaries of cognitive

8 PHILIP ROBBINS A N D MURAT AYDEDE

systems lie outside the envelope of individ-ual organisms, encompassing features of the physical and social environment (Clark & Chalmers, 1998; Wilson, 2004). In this view, the mind leaks out into t he world, and cog-nitive activity is distributed across individ-uals and situations. This is not your grand-mother's metaphysics of mind; this is a brave new world. Why should anyone believe in it?

One part of the answer lies in the promise of dynamical systems theory - the intel-lectual offspring of classical control theory, or cybernetics (Ashby, 1956; Wiener, 1948; Young, 1964) - as an approach to model-ing cognition (Beer, 1995; Thelen & Smith, 1994; van Gelder, 1995). Using the tools of dynamical systems theory, one can describe in a mathematically precise way how various states of a cognitive system change in rela-tion to one another over time. Because those state changes depend as much on changes in the external environment as on changes in the internal one, it becomes as important for cognitive modeling to track causa] pro-cesses that cross the boundary of the indi-vidual organism as it is t o track those that lie within that boundary. In short, insofar as the mind is a dynamical system, it is natu-ral to think of it as extending not just into the body but also into the world. The result is a radical challenge to traditional ways of thinking about the mind, Cartesian internal-ism in particular:

The Cartesian tradition is mistaken in sup-posing that the mind is an inner entity of any kind, whether mind-stuff, brain states, or whatever. Ontobgically, mind is much more a matter of what we do within environmental and social possibil-ities and bounds. Twentieth-century anti-Cartesianism thus draws much of mind out, and in particular outside the skull, (van Gelder, 1995, p. 380}

Implicit in this passage is a kind of slippery slope argument premised on a broad theo-retical assumption. Grant that cognition is embodied and embedded - something that the dynamical systems approach takes more or less as a given - and it is a short distance

to the conclusion that cognition is extended as well. Or so the reasoning goes.

Another part of the motivation behind the extension thesis traces back to a fic-tional (but realistic) scenario that Clark and Chalmers (1998) describe. They introduce a pair of characters named Otto and Inga. Ot to is an Alzheimer's patient who supple-ments his deteriorating memory by carry-ing around a notebook stocked with use-ful information. Unable to recall the address of a museum he wishes to visit, Ot to pulls out his trusty notebook, flips to the rele-vant page, looks up the address, and pro-ceeds on his way. Neurotvpical Inga, in con-trast, has an intact memory and no need for such contrivances. When she decides to visit the museum, she simply recalls the address and sets off. Now, there are clear differences between the case of Ot to and the case of Inga; Ot to stores the information externally (on paper), whereas Inga stores it internally (in neurons); Ot to retrieves the information by visual lookup, whereas Inga uses some-thing like introspective recall; and so on. But according to Clark and Chalmers, these differences are relatively superficial. What is most salient about the cases of Ot to and of Inga, viewed through a functionalist lens, are the similarities. Once these similarities are given their due, the moral of the story becomes clear: "When it comes to belief, there is nothing sacred about skull and skin. What makes some information count as a belief is the role it plays, and there is no rea-son why the relevant role can be played only from inside the body" (Clark & Chalmers, 1998, p. 14). As for the fact that this con-ception of mind runs afoul of folk intu-itions, well, so much the worse for those intuitions.

This conclusion is not forced on us, how-ever, and a number of theorists have urged that we resist it. For example, Rupert (2004) argues that generalizing memory to include cases like Otto 's would have the untoward effect of voiding the most basic lawlike gen-eralizations uncovered by traditional mem-ory research, such as primacy, recency, and interference effects - and without furnishing anything comparably robust to substitute in

A SHORT PRIMER O N SITUATED COGNITION 9

their place. In short, insofar as the goal of scientific inquiry is to carve nature at its joints, and lawlike regularities are the best guide to the location of those joints, it is not clear that a fruitful science of extended memory is possible, even in principle. More generally, Adams and Aizawa (2008) con-tend that the standard argument for pushing the boundary of cognition beyond the indi-vidual organism rests on conflating the meta-physically important distinction between causation and constitution. As they point out, it is one thing to say that cognitive activity involves systematic causal interac-tion with things outside the head, and it is quite another to say that those things instan-tiate cognitive properties or undergo cogni-tive processes. Bridging this conceptual gap remains a major challenge for defenders of the extended mind.

4. Coda

Situated cognition is a many-splendored enterprise, spanning a wide range of projects in philosophy, psychology, neuroscience, anthropology, robotics, and other fields. In this chapter we have touched on a few of the themes running through this research, in an effort to convey some sense of what situ-ated cognition is and what the excitement is about. The twenty-five chapters that follow it develop these themes, and other themes in the vicinity, in depth. Both individually and collectively, these chapters reveal what "getting situated" means to cognitive sci-ence, and why it matters.

References

Adams, F., & Aizawa, K. (2008). The bounds of cognition. Oxford: Blackwell.

Anderson, M. L. (2003). Embodied cognition: A field guide. Artificial Intelligence, 149, 91-130.

Ashby, W. R. (1956). Introduction to cybernetics. New York: Wiley.

Beer, R. D. (1995). A dynamical systems perspec-tive on agent-environment interaction. Artifi-cial Intelligence, 72,173-215.

Block, N. (2005). Review of Alva No's Action in Perception. Journal of Philosophy, 102, 259-272.

Brooks, R. (1991). Intelligence without represen-tation. Artificial Intelligence, 47,139-159.

Cacioppo, J. T., Priester, J. R., & Bemston, G. G. (1993). Rudimentary determination of atti-tudes: II. Arm flexion and extension have dif-ferential effects on attitudes. Journal of Person-ality and Social Psychology, 65, 5-17.

Chen, S., & Bargh, J. A. (1999). Consequences of automatic evaluation: Immediate behavior dispositions to approach or avoid the stimulus. Personality and Social Psychology Bulletin, 25, 215-224.

Clancey, W. ]. (1997). Situated cognition. Cam-bridge: Cambridge University Press.

Clark, A. (1989). Microcognition. Cambridge, MA: MIT Press.

Clark, A. (1997). Being there. Cambridge, MA: MIT Press.

Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58,10-23.

Gallagher, S. (2005). How the body shapes the mind. Oxford: Oxford University Press.

Gibson, J. J. (1979). The ecological approach to visual perception. Boston: Houghton Mifflin.

Gigerenzer, G. (2000). Adaptive thinking. Oxford: Oxford University Press.

Glenberg, A. M., & Kaschak, M. P. (2002). Grounding language in action. Psychonomic Bulletin and Review, 9, 558-565.

Harnad, S. (1990). The symbol grounding prob-lem. Physica D, 42, 335-346.

Hurley, S. L. (1998)- Consciousness in action. Cambridge, MA: Harvard University Press.

Hutchins, E. (1995). Cognition in the wild. Cam-bridge, MA: MIT Press.

Kirsh, D. (1995). The intelligent use of space. Arti-ficial Intelligence, 7, 31-68.

Kirsh, D., & Maglio, P. (1994). On distinguish-ing epistemic from pragmatic action. Cognitive Science, 18, 513-549.

Lakoff, G., & Johnson, M. (1999]. Philosophy in the flesh. New York: Basic Books.

Markman, A. B., 81 Dietrich, E. (21300). Extending the classical view of representation. Trends in Cognitive Sciences, 4, 470-475.

Nagel, T. (1974). What is it like to be a bat? Phil-osophical Review, 82, 435-450.

Niedenthal, P. M., Barsalou, L. W., Winkiel-man, P., Krauth-Gruber, S., 8t Ric, F. (2005). Embodiment in attitudes, social perception, and emotion. Personality and Social Psychology Review, 9,184-211.

10 PHILIP ROBBINS A N D MURAT AYDEDE

No, A. (2004). Action in perception. Cambridge, MA: MIT Press.

O'Regan, J. K. (1992). Solving the "real" mysteries of visual perception: The world as an outside memory. Canadian Journal of Psychology, 46, 461-488.

Robbins, P. (20083. Consciousness and the social mind. Cognitive Systems Research, 9, 15-

Rupert, R. (2004]. Challenging the hypothesis of extended cognition. Journal of Philosophy, 101, 389-428.

Thelen, E., & Smith, L. B. (1994). A dynamic systems approach to the development of cog-nition and action. Cambridge, MA: MIT Press.

van Gelder, T. (1995). What might cognition be, if not computation? Journal of Philosophy, 91, 545-381.

Vera, A. H., & Simon, H. A. (1993). Situated action: A symbolic interpretation. Cognitive Science, 17, 7-48.

Wiener, N. (1948). Cybernetics; or, Control and communication in the animal and the machine. New York: Wiley.

Wilson, M. (2002). Six views of embodied cogni-tion. Psychonomic Bulletin and Review, 9, 625-636.

Wilson, R. A. (2004). Boundaries of the mind. Cambridge: Cambridge University Press.

Young, J. Z. (1964). A model of the brain. Oxford: Oxford University Press.

I;- . r I -

fsi C H A P T E R 17

P; [ Situating Rationality I"

Ecologically Rational Decision Making i I with Simple Heuristics

Henry Brighton and Peter M. Todd

i : & i !,JLM i

rafe j

I

N f-; i.

i. Introduction

Many within cognitive science believe that rational principles rooted in probability the-ory and logic provide valuable insight into the cognitive systems of humans and ani-mals. More than this, some say that ratio-nal principles, such as Bayesianism, algorith-mic information theory, and logic, not only provide elegant and formally well under-stood frameworks for thinking about cog-nition but also are the very principles gov-erning thought itself, guiding inferences and decisions about the world (e.g., Chater & Vitnyi, 2003; Feldman, 2003). It is not diffi-cult to see why ascribing such principles to the cognitive system is a tempting and desir-able goal: if correct, these principles would provide universal normative laws governing the cognitive system. Even faced with the diversity of tasks and environments handled by the cognitive system, a valid universal principle would have the advantage of pro-viding a theoretical handle with which to grasp the range of cognition in a unified man-ner. In contrast, situated theories of cogni-

tion take the specific and concrete details of the interaction between mind and envi-ronment to matter significantly, so much so that seeking universal principles of ratio-nality is seen as one of the primary wrong turns in the path of traditional approaches to understanding cognition (e.g., Smith, 999)-

Other approaches to cognition find sim-ilar fault with the search for univer-sal mechanisms or all-powerful inferential machinery. The recently developed view of ecological rationality places a strong focus on the structural properties of environments, and takes a structure-specific, and, as we will show in this chapter, situated approach to the study of cognitive processes. Ecological rationality emerges when simple, cognitively plausible heuristic processes exploit spe-cific environmental characteristics to make adaptive choices in particular circumstances (Gigerenzer, Todd, & ABC Research Group, 1999). Adaptive choices are those that help an organism function successfully in its envi-ronment. Rather than view the mind as a general-purpose processor endowed with

3"

SITUATING RATIONALITY 3*3

the considerable computational resources demanded by classical notions of rational inference, we instead explore the possibility that the mind is more like a mixed bag of cognitive processes, each working within the limitations of the cognitive system, and each tuned to specific structures occurring in natural environments. We term this mixed bag of simple heuristics the adaptive toolbox (e.g., Gigerenzer & Selten, 2001). In section 3 we firm up this metaphor by providing examples of simple heuristics, and we dis-cuss work that demonstrates, through com-putational modeling and experimental stud-ies, that these heuristics are both plausible and often more effective than traditional models of inference.

In section 4 we address our main con-cern: situating rationality in such a way that the specific, concrete, engaged, and located nature of human cognition is given due sig-nificance. We argue that classical visions of rationality fail to consider the compu-tational limitations of the cognitive system and the importance of the role of cognitive exploitation of environment structure. On this view, normative theories derived from classical notions of rationality prove limited in their ability to capture the specific nature of how the cognitive system interacts adap-tively with the environment in inference and decision-making tasks. Although systems of logic, for example, can be tailored to spe-cific tasks and environmental contexts (e.g., Stenning & van Lambalgen, 2004), the issue of how the cognitive system might plau-sibly process information to achieve logi-cal consistency, and whether such a task is computationally feasible, remains unclear. In contrast, we argue that ecological ratio-nality bounds rationality to the world by considering both ecological context and con-straints on cognitive processing (Todd & Gigerenzer, 2003). This view of rationality is closely allied with the situated cognition movement, and in section 5 we clarify the ways in which ecological rationality can be understood in terms of the key dimensions of situated approaches proposed by Smith (999)-

2. Mind and Environment: The Terms of the Relationship

In what ways does the environment influ-ence the contents of the mind? A wide range of possibilities exists. For some aspects of the cognitive system it could be that the envi-ronment plays an insignificant role in influ-encing the contents of the mind. The human language faculty, for instance, is thought by some researchers to be shaped not by the environment but by boundary conditions internal to the mind, in the sense that the linguistic system is an optimal solution for meeting the requirements of transforming thought between the conceptual-intentional system and the articulatory-motor system (Chomsky, 1995). According to this view, if we want to understand the language faculty, then it pays not to consider the environ-ment at all, either over evolutionary or onto-genetic timescales. But for many aspects of the cognitive system we cannot ignore the important impact of the environment, and several possible forms of relationship between mind and environment need to be considered. Within psychology, the ecolog-ical approach to cognition seeks to under-stand the terms of such relationships. We begin by considering three metaphors (Todd & Gigerenzer, 2001) that aim to capture some key relationships.

2.1. Three Kinds of Relationships

To understand the contents of the mind, we should consider the environment in which it acts and in which it has evolved. This ecolog-ical, situated perspective has been promoted within cognitive psychology in particular by Roger Shepard's work (see, e.g., Shepard, 2001, and other papers in the special issue of Behavioral and Brain Sciences on Shepard's research and related efforts it has inspired), focusing on a particular vision of how the external world shapes our mental mecha-nisms. For Shepard (2001), much of per-ception and cognition is done with mirrors: key aspects of the environment are inter-nalized in the brain "by natural selection

HENRY BRIGHTON A N D PETER M. T O D D

specifically to provide a veridical represen-tation of significant objects and events in the external world" (p. 2). In particular, Shepard considers how the cognitive system reflects features of the world, or laws, which sup-port, for example, tracking the movement of objects in Euclidean space. These abstract principles are based "as much (or possibly more) in geometry, probability, and group theory, as in specific, physical facts about concrete, material objects" (Shepard, 2001, p. 601). Thus, the cognitive system possesses deeply internalized, abstract, and univer-sal reflections of physical reality. Shepard's work can be viewed as ecological in that an understanding of mind is taken to be fundamentally dependent on identifying the important properties of the physical world. Yet this view also turns on the contentious issue of the mind representing properties of the world, albeit at a very abstract level. Without entering into arguments over the need for representations of any sort (for a discussion of the key issues, see, e.g., Brooks, 1991; Markman & Dietrich, 2000), we can still question whether assumed representations should be veridical, constructed to accu-rately reflect the world, or instead be useful in an adaptive sense. In short, whatever the functional role of such representations, this mirrorlike relationship characterizes those properties of mind that are present as a result of an internalization of universal laws governing physical environments.

A less exacting view of internalization can be seen in the work of Egon Brunswik (1955], who proposed a lens model that reconstructs a representation of a distal stimulus on the basis of the current proximal cues (whose availability could vary from one decision sit-uation to the next) along with stored knowl-edge of the environmental relationships between those perceived cues and the stim-ulus. These relationships were later usually conceived of as correlations in the field of social judgment theory that followed f rom Brunswik's work (see Hammond & Stewart, 2001). For Brunswik, the mind models and projects the world more than reflects it (or, as he also put it, mind and world accommo-date each other like husband and wife).

Herbert Simon (1990) expressed a still looser coupling between mind and envi-ronment: bounded rationality, he said, was shaped by a pair of scissors whose two blades are the characteristics of the task environment and the computational capa-bilities of the decision maker. Here, com-putational capabilities refer to sensory, neu-ral, and other mental characteristics that may impose cognitive limitations on, for example, memory and processing. Crucially, these capabilities, when coupled with cer-tain characteristics of the environment, can complement one another. Rather than the mind reflecting or projecting properties of the environment, Simon's scissors metaphor highlights a very different kind of relation-ship in which properties of mind are viewed as fitting properties of environments in an exploitative and complementary relation-ship. From this perspective, it is less clear how meaningfully one can characterize the relationship between mind and environment in terms of internalization or representa-tion, as some properties of the mind can be only implicitly related to the environ-ment rather than more direcdy, as suggested by metaphors of mirror images or projec-tions. Considering this kind of exploita-tive relationship led Simon (1956) to con-sider, in the context of decision making, "how simple a set of choice mechanisms we can postulate and still obtain the gross features of observed adaptive choice behav-ior" (p. 129). This question highlights how an exploitative relationship between mind and environment has implications for the kind of cognitive machinery used by the mind: as Simon and others have since shown, simple, boundedly rational decision mecha-nisms coupled with the right environmental context can yield adaptive choice behavior that is typically attributed to more complex and information-hungry mechanisms.

2.2. Appropriate Metaphors for Higher-Level Cognition

We expect that the mind draws on mecha-nisms akin to all three tools, mirrors, lenses, and scissors, from its adaptive toolbox

SITUATING RATIONALITY 325

(Gigerenzer & Todd, 1999). The question now becomes, Where can each be used, or where does each different view of sit-uated cognition best apply? In perception, using Shepard's mirror or Brunswik's lens may often be the right way to look at things, but there are also examples in which these tools are inappropriate. Consider the prob-lem of a fielder trying to catch a ball coming down in front of her. The final destination of the ball will be complexly determined by its initial velocity, its spin, the cffects of wind all along its path, and other causal factors. But rather than perceive all these character-istics, reflect or model the world, and com-pute an interception point to aim at, the fielder can use a simple heuristic: fixate on the ball and adjust her speed while running toward it so that her angle of gaze - the angle between the ball and the ground from her eye - remains constant (McLeod & Dienes, 1996). By using this simple gaze heuristic, the fielder will catch the ball while running. No veridical representations or models of the world are needed - just a mechanism that fits to and exploits the relevant structure of the environment; namely, the single cue of gaze angle. How widely such scissors-like heuris-tics can be found in perception remains to be seen, bu t some (e.g., Ramachandran, 1990) expect that perception is a bag of tricks like this rather than a box of mirrors.

When we come to higher-order cogni-tion and decision making, our main con-cern in this chapter, Simon's cutting per-spective seems the most appropriate way to extend Shepard's and Brunswik's eco-logical views. Consider a simple decision rule tha t has been proposed as a model of human choice: the Take the Best heuristic (Gigerenzer & Goldstein, 1996). To choose between two options on the basis of several cues known about each option, this heuris-tic says to consider one cue at a t ime in order of the ecological validity of each (i.e., how often each cue makes correct decisions), and to stop this cue search with the first one that distinguishes between the options and make the final decision using only that cue (we will explain this heuristic in more detail later in the chapter). This "fast and frugal"

heuristic makes decisions about as well as multiple regression in many environments (Czerlinski, Gigerenzer, & Goldstein, 1999) but usually uses far less information (cues) in reaching a decision. It does not incor-porate enough knowledge to reasonably be said to reflect the environment, nor even to model it in Brunswik's sense (because it only knows cue order, not exact validities), but it can certainly match and exploit envi-ronment structure: when cue importance is distributed in something like an expo-nentially decreasing manner (as is the case in some environments), Take the Best per-forms about as well on training data sets as multiple regression or any other linear deci-sion rule (Martignon & Hoffrage, 1999) and generalizes to new data sets even better.

As another example, the QuickEst heuristic for estimating quantities (Hertwig, Hoffrage, & Martignon, 1999) is similarly designed to use only those cues necessary to reach a reasonable inference. QuickEst makes accurate estimates with a minimum of information when the criteria of the objects it operates on follow a J-shaped (power law) distribution, such as the sizes of cities or the number of publications per psychologist. Again this crucial aspect of environment structure is nowhere built into the decision mechanism, but by processing the most important cues in an appropriate order, QuickEst can exploit that structure to great advantage. Neither of these heuristics (which we will describe in more detail in the next section) embodies logical rationality -they do not even consider all the available information - bu t rather they demonstrate ecological rationality, making adaptive deci-sions by relying on the structure of the envi-ronment.

Why might it be that Simon's scissors could help us understand cognitive mecha-nisms more than Shepard's mirror? We (and others) suspect that humans often use sim-ple decision-making mechanisms that are built on (and receive their inputs from) much more complex lower-level percep-tual mechanisms. For instance, the recogni-tion heuristic, an elementary mechanism for deciding between two options on the basis of

3 * 6 HENRY BRIGHTON A N D PETER M. T O D D

M.

I f I ft* i i'jb -:4 -1'0 t

t

if

which of them are merely recognized, sim-ply uses the binary cue of recognized ver-sus not recognized; however, the computa-tional machinery involved in the lower-level assessment of whether a voice, or face, or name actually is recognized involves con-siderable complexity (Todd, 1999). If these decision heuristics achieve their simplicity in part by minimizing the amount of infor-mation they use, then they are less likely to reflect the external world and more likely to exploit just the important, useful aspects of it, as calculated and distilled by the percep-tual system (which may well base its compu-tations on a more reflective representation).

Thus, in extending the search for the imprint of the world on the mind from per-ception to higher-order cognition, we should probably look less for reflections and more for complementary pairings la Simon's two scissors blades. This approach to studying environmentally situated deci-sion mechanisms is just what we shall intro-duce in this chapter. While Simon studied bounded rationality, we use the term ecolog-ical rationality to emphasize the importance of the match between the structure of infor-mation in the environment and the structure of information processing in the mind. In the next section, we introduce the notion of the adaptive toolbox and describe how its con-tents can be studied, before expanding on some examples of its contents by describing simple ecologically rational decision heuris-tics. In section 4 we develop further the con-cept of ecological rationality by discussing how it contrasts with traditional notions of rationality, and why, in the context of the study of mind, ecological rationality is a more appropriate notion when considering aspects of high-level cognition. In section 5 we relate these discussions to the study of situated cognition by framing ecological rationality as a form of situated cognition.

3. The Adaptive Toolbox and Its Contents

It is certainly not the case that all of human behavior is ecologically rational, as defined

here. For instance, people can (if given the luxury of sufficient time and training) use more general methods of reasoning accord-ing to traditional norms of rationality, such as the tools of logic or probability theory, to come to decisions with little concern for adapting their reasoning to the spe-cific structure of the current task environ-ment. Or people may use simple decision heuristics that try to exploit some features of the environment to allow for cognitive shortcuts, but in the wrong environment, so that biased decision making arises (as studied in the heuristics and biases research tradition; see, e.g., Kahneman, Slovic, & Tversky, 1982). But we propose that much of human decision making is ecologically rational, guided by typically simple deci-sion heuristics that exploit the available structure of the environment to make good choices. Given the right environmental cir-cumstances, these simple methods perform adequately for many tasks and sometimes better than more complex mechanisms. One consequence of this theory is that a single all-purpose decision-making system is no longer the appropriate unit of study, as different tasks call for different simple mechanisms. The idea of the adaptive tool-box leads us to consider a collection of sim-ple mechanisms drawn on by the cogni-tive system. We view these mechanisms as structure specific rather than domain spe-cific. In contrast to the concept of domain specificity, structure specificity is the ability of a process to deal effectively with infor-mational structures found in environments that may or may not be encountered in multiple domains (e.g., a systematic corre-lation between recognition knowledge and some criterion of interest, which we discuss herein). These mechanisms are built from basic, cognitively primitive building blocks for information search, stopping the search, and making a decision based on the search's results. How heuristics are constructed using these building blocks, when and why one heuristic is used over another, and how and how well they work in different situations are all key issues confronting this research program. In this section, we give a taste

SITUATING RATIONALITY V-1

of how these issues are addressed and dis-cuss how this approach can b e a productive route to understanding the cognitive system in terms of simple process models tuned to environment structures.

A number of steps are involved in study-ing the contents of the adaptive toolbox. First, after identifying a particular eco-logically important decision domain, we must determine the structure of informa-tion available to people in that domain. As Shepard (2001) indicated, this involves dis-covering the "general properties that char-acterize the environments in which organ-isms with advanced visual and locomotor capabilities are likely to survive and repro-duce" (p. 581) - these might include power laws governing scale invariance (Bak, 1996; Chater & Brown, 1999) or costs of t ime and energy in seeking information (Todd, 2001). Shepard (1987) considers only the longest-applying physical laws as stable parts of the environment, avoiding discussion of the bio-logical and social realms that he feels have not been around as long and so will not have exerted as much pressure on our and other animals' evolved cognition. However, we have certainly evolved adaptive responses to these realms of challenges as well, and so we should extend the study of environmentally matched decision heuristics to biological and social domains (as is done within evolution-ary psychology; e.g., Barkow, Cosmides, & Tooby, 1992; Buss, 2005). This means we should look for environmental-information regularities tha t may be internalized to guide our cognition when situated in different evolutionarily important domains, such as predator risk (e.g., Barrett, 2005), knowledge of infection and disease transmission (e.g., Rozin & Fallon, 1987), understanding genetic dynamics (e.g., kin selection; Hamilton, 1996), and social interactions of various types (e.g., on the dynamics of signaling between agents with conflicting interests, see Zahavi & Zahavi, 1997; on mate choice, social exchange, and dominance hierarchies, see Buss, 2005).

With these characteristic environment structures in mind as one half of Simon's scissors, we can look more effectively for

the decision mechanisms that form the other matching half. This can involve a further set of steps, as follows:

1. Investigating, through simulation, can-didate models of cognitive processes built from elementary processing abil-ities such as search and recognition; to achieve this, a model selection criterion is required (e.g., performance criteria such as predictive accuracy, frugality of information use). These yardsticks set the scene for comparisons with other models.

2. Identifying if and when these heuris-tics perform well in certain environ-ments, and what the characteristics of these environments are, often using ana-lytic methods. Is it possible to give pre-cise environment-structure conditions for good or poor performance of the heuristics? Do these conditions match those of typical environments inhabited by humans?

3. With an understanding of how the model and the environment (or task) structure match, carrying out empirical studies t o see if humans use these pro-cesses in appropriate situations.

The use of elementary processing abili-ties to guide the construction of candidate models reflects a commitment to bot tom-up design. One consequence of this approach is that it leads to simple and testable models with few parameters and therefore makes for a more transparent relation between the-ory and data. Another consequence, which affects a core concern for the study of ecological rationality, is that the resultant cognitive models, by virtue of their sim-plicity and close reliance on fundamental processing abilities, are more likely to be cognitively plausible. This approach adopts the view that functioning cognitive models can be built with nontrivial consequences without needing to be monolithic, gener-ally applicable, and computationally com-plex. In short, robust cognitive processing is achieved with simple and ecologically targeted mental heuristics. The concept of

32S HENRY BRIGHTON A N D PETER M. T O D D

ecological rationality and the adaptive tool-box, in some respects, is close in spirit to Anderson's rational analysis, where a con-sideration of the structure of the environ-ment constrains the development of cogni-tive models, and a focus on the plausibility of the cognitive models then steers future development (Anderson, 1990; Oaksford & Chater, 1998). However, rational analysis, in contrast to the approach explored here, places less of an emphasis on bot tom-up design. One consequence of this difference is that the adaptive toolbox leads to a con-sideration of multiple simple models rather than fewer, more complex models.

What is in the adaptive toolbox? Several classes of simple heuristics for making differ-ent types of decisions in a variety of domains have been investigated (see, e.g., Gigerenzer et al., 1999; Kahneman et a l , 1982; Payne, Bettman, & Johnson, 1995; Simon, 1990), including ignorance-based heuristics that make decisions based on a systematic lack of information, one-reason heuristics that make a choice as soon as a single reason is found on which to base that choice, elim-ination heuristics that whittle down a set of choices using as few pieces of informa-tion as possible until a single choice is deter-mined, and satisficing heuristics that search through a sequence of options until a good-enough possibility is found. Other tools are also present, and more await discovery. Here we present three examples of the heuristic tools in the toolbox.

3.1. Paired Comparison Using the Recognition Heuristic

The capacity for recognition is common to many species. The most basic cogni-tive heuristic we will focus on exploits this capacity to make inductive inferences. Given the task of deciding which of two objects in the world scores higher on some criterion of interest, the recognition heuris-tic can provide a quick and robust decision procedure by exploiting a lack of knowl-edge. If one of the objects being considered is recognized and the other is not, then the recognition heuristic tells us to judge the

recognized object as scoring higher on the criterion. For example, given the names of two tennis players, the recognition heuris-tic simplifies the task of deciding which of these two tennis players is most likely to win the next Grand Slam tournament: if we only recognize one of the players and not the other, then the recognition heuristic tells us to pick the player we have heard of (Pachur & Biele, 2007). Similarly, given the task of choosing which of two cities has a higher population, the recognition heuristic tells us to pick the city we have heard of over the one we have not (Goldstein & Gigeren-zer, 1999, 2002). Clearly, the appropriate use of this heuristic depends on (a) applicabil-ity, because to apply the recognition heuris-tic certain conditions must be met, and (b) validity, because the ability to apply the heuristic does not necessarily imply that it will lead to accurate inferences. Specifically, the recognition heuristic is applicable only when the decision maker has an intermedi-ate amount of (recognition) knowledge, not complete knowledge or complete ignorance (which would render the recognition heuris-tic unusable). And this heuristic is only valid when the partial recognition knowledge is systematic; that is, correlated with the crite-rion on which the decision is being made -for example, given that winning tennis play-ers are talked about in person and in media more than losing ones, more people will rec-ognize the best players, meaning that recog-nition is correlated with past success, and hence, presumably, with future success as well. Both these conditions place restrictions on the kinds of environments in which the recognition heuristic will perform well, and defining such environments and determin-ing how the heuristic operates in them are precisely the sorts of questions addressed and explored by the study of ecological rationality.

Using simple heuristics can also lead to surprising outcomes, which the eco-logical rationality framework can predict and explain. For example, when knowledge about and recognition of the objects in the environment (e.g., cities) is positively cor-related with the criterion of interest (e.g.,

SITUATINC RATIONALITY 329

population), then the recognition heuristic can lead an individual with less knowledge to make more reliable decisions than a per-son with more knowledge. This less-is-more effect has been confirmed in simulations and demonstrated in experiments involving individuals (Goldstein & Gigerenzer, 1999, 2002) and groups of subjects (Reimer & Katsikopoulos, 2004). This example is strik-ing because it shows how a simple mech-anism built on a basic cognitive capacity, here recognition memory, when used in the right environmental setting, can enable the cognitive system to exploit environmen-tal structure and subsequently make good decisions with little information or pro-cessing.

3.2. Paired Comparisons Using Take the Best

If only one object in a paired compari-son task is recognized, then there is little choice but to apply the recognition heuris-tic. But when both objects are recognized, and knowledge of several cues about each object is available to aid the decision, then many possible decision processes exist. For example, the paired comparison task is a special case of the general task of learn-ing to categorize f rom labeled examples, which is explored thoroughly in machine-learning research (Mitchell, 1997). In that field, a litany of potential processes exist, typically complex algorithms designed from an engineering perspective to approximate a general solution to the problem of learn-ing from examples, which leads these meth-ods to veer from considerations of cognitive plausibility. In contrast, the study of ecolog-ical rationality from the perspective of the adaptive toolbox takes the issues of cogni-tive plausibility and the specific nature of the task as fundamental. These emphases are reflected in a bot tom-up approach in which simple processes are built from ele-mentary building blocks chosen to match a particular task environment. In particular, as mentioned earlier, Gigerenzer and col-leagues (1999) have explored the following types of building blocks for processing cues

representing features of objects encountered in the world:

1. Search rules, which define how informa-tion in the form of cues is searched for. For example, one possible search rule is to look through cues in an order that reflects how useful these cues have been in the past.

2. Stopping rules, which define when cue search is to be terminated. For example, given the task of comparing two objects in terms of their cue values, search can be terminated when a stopping crite-rion of different cue values for the two objects is met.

3. Decision rules, which define how the information found by the first two building blocks is used to make a deci-sion. For example, given information about a cue that differs in value for two objects, the object with the higher cue value could be chosen.

Take the Best is a simple heuristic built from three such building blocks where (a) cues are searched in order of their ecologi-cal validity, (b) search stops at the first dis-criminating cue (i.e., the first cue that has a different value for each object, and hence discriminates between the two objects), and (c) the object selected is the one indicated by the discriminating cue. Ecological valid-ity is a property of a cue, which indicates how frequently in the past the discrimi-nating cue picked out the object with the higher criterion value. (Discrimination rate is another important property of cues, indi-cating how often a given cue discriminates between pairs of objects in some environ-ment.) For example, Take the Best could decide which of two tennis players is more likely to win an upcoming competition by first considering the most valid cue, say, "Has this player won a Grand Slam competition in the past?" If this cue discriminates - that is, it is true for one player and not the other - then Take the Best will stop information search and select the previous-winning player over the other. If t he first cue does not discrim-inate, then fur ther cues are considered in

53 HENRY BRIGHTON A N D PETER M. T O D D

order of ecological validity until one is found that does discriminate and is then used by itself to determine the decision.

Unlike many other models of decision making, which typically take all cues into consideration and combine them somehow to yield a decision, Take the Best is fru-gal in its use of information. The decision is made only on the basis of the first dis-criminating cue, and all other information is ignored. In this sense, Take the Best employs one-reason decision making. But does Take the Best suffer, in terms of performance, by ignoring so much of the available informa-tion? No - Take the Best often performs just as well as other less frugal and more compu-tationally intensive models such as multiple linear regression, even though it uses only a fraction of the available information. But even more surprising is the fact that in a con-siderable number of decision environments examined so far, Take the Best can outper-form rival models of decision making when generalizing to new decisions (Czerlinski et al., 1999). Furthermore, recent work demonstrates that Take the Best can even beat the key models of inductive inference used in machine-learning research on con-nectionist, rule-based, and exemplar-based approaches - as long as the environment has a particular information structure (Brighton, 2006; Chater, Oaksford, Nakisa, & Reding-ton, 2003). Thus, this work has shown in principle that decision processes built from simple, cognitively plausible building blocks that use little processing and little infor-mation in ways that are matched to the task environments in which they are sit-uated can outperform some of the most widely used and studied models of induction that take a domain-general, environment-agnostic approach.

But in practice, do people use such sim-ple, fast-and-frugal heuristics to make deci-sions? A growing body of experimental and empirical work is demonstrating that peo-ple do use one-reason decision heuristics in appropriately structured environments, such as where cues act individually to sig-nal the correct response (Rieskamp & Otto, 7005) or where information is costly or time

consuming to acquire (Brder, 2000; Brder & Schiffer, 2003; Newell & Shanks, 2003-Rieskamp & Hoffrage, 1999). People also make socially and culturally influenced deci-sions based on a single reason through imi-tation (e.g., in food choice; Ariely & Levav 2000), norm following, and employing pro-tected values (e.g., moral codes that admit no compromise, such as never taking an action that results in human death; see Tan-ner & Medin, 2004).

3.3. Estimation Using QuickEst

Not all choices in life are presented to us as convenient pairs of options, of course. Often we must choose between several alternatives, such as which restaurant to go to or which habitat to settle in. In situa-tions where each available cue dimension has fewer values than the number of avail-able alternatives, one-reason decision mak-ing will usually not suffice, because a sin-gle cue will be unable to distinguish among all of the alternatives. For instance, knowing whether each of fifteen cities has a river is not enough information to decide which city is most habitable. But this does not doom the fast-and-frugal reasoner to a long pro-cess of cue search and combination in these situations. Again, a simple stopping rule can work to limit information search: seek cues (in an order specified by the search rule) only until enough is known to make a decision. But now a different type of decision rule is needed instead of relying on one reason. One way to select a single option from among multiple alternatives is to follow the sim-ple principle of elimination (Tversky, 1972): successive cues are used to eliminate more and more alternatives and thereby reduce the set of remaining options, until a single option can be decided on.

The QuickEst heuristic (Hertwig et al., 1999) is designed to estimate the values of objects along some criterion while using as little information as possible. The estimates are constrained to map onto certain round numbers (e.g., when estimating city popu-lation sizes, QuickEst can return values of 100,000,150,000, 200,000, 300,000, and other

SITUATING RATIONALITY 351

similarly round numbers), so this heuristic can be seen as choosing one value from sev-eral possibilities. The elimination-based esti-mation process operates like a coal sorter, in which chunks of coal of various sizes first foil over a small slit, through which the smallest pieces fall into a bin for fine-grained coal; the bigger pieces that remain then roll over a wider slit that captures medium--sized pieces of coal into a medium bin; and finally the biggest chunks roll into a large coal bin. In the case of QuickEst applied to city population estimates, the coal chunks are cities of different population sizes; the bins are the rounded-number size estimates; and the slots are cues associated with city size, ordered according to the average size of the cities without that cue (e.g., because most small cities do not have a professional sports team, this could be one of the first cues checked - i.e., one of the first slots that the cities roll past). To estimate a city's size, the QuickEst heuristic looks through the cues or features in order until it comes to the first one that the city does not pos-sess, at which point it stops searching for any further information (e.g., if a city pos-sesses the first several features in order but lacks a convention center, the city falls into that bin and search will stop on that cue). QuickEst then gives the rounded mean cri-terion value associated with the absence of that cue as its final estimate (e.g., the mean size of all entries in the bin for cities without an exposition site). Thus, in effect, QuickEst uses features that are present to eliminate all smaller criterion categories, and absent features to eliminate all larger criterion cat-egories, so that only one criterion estimate remains. No cue combination is necessary, and no adjustment from further cues is pos-sible.

QuickEst proves to be fast and frugal, as well as accurate, in environments character-ized by a distribution of criterion values in which small values are common and big val-ues are rare (a so-called J-shaped distribu-tion, where the J is seen on its side). Such distributions characterize a variety of natu-rally occurring phenomena, including many formed by accretionary growth. This growth

pattern applies to cities (Makse, Havlin, & Stanley, 1995), and indeed big cities are much less common than small ones. As a consequence, when applied to a data set of cities, QuickEst is able to estimate rapidly the small sizes that most of them have.

4. Rationality in the Real World: From the Classical to the Ecological

The heuristics introduced in the previous section are tools for making inductive infer-ences. Induction is the task of using the past to predict and make decisions about the future. Without being able to look into the future, we can never answer with certainty questions such as, say, which tennis player will next win Wimbledon. W e instead have to make an inductive inference, a predic-tion about the future, based on observations and knowledge we have acquired in the past. Clearly, by using prior experience we can often make better than chance predictions about future events in the world. This would not be possible if the world were unstruc-tured and behaved randomly. Fortunately, most environments we face are highly struc-tured, so that principled decision making serves us well. But what principles should guide our decision making? Many theories take what we will term classical rationality as a source of answers. This is to say that for the task of reasoning under uncertainty, clear normative principles such as Bayesian infer-ence and variants of Occam's razor based on algorithmic information theory tell us what the rational course of action is (e.g., Chater, 1999; Hutter, 2005; Pearl, 1988). Yet humans often deviate from the classically rational: we make errors in a sometimes-systematic fashion and appear to only partially adhere to normative ideals. This view of occasion-ally error-prone human behavior is most forcefully argued within the heuristics-and-biases tradition, which proposes that peo-ple rely on a limited number of heuristic principles that reduce the complex tasks of assessing probabilities and predicting values to simpler judgmental operations. In gen-eral, these heuristics are quite useful, but

3 3 J HENRY BRIGHTON A N D PETER M. T O D D

sometimes they lead to severe and system-atic errors (Tversky, 1972, p. 1124).

In this section we develop and justify why the classical view of rationality, when adopted as a concept with which to under-stand human decision making and infer-ence, fails to capture significant aspects of the inference task. We will contrast classi-cal rationality with ecological rationality and argue that the latter offers a far more pro-ductive concept with which to understand human decision making, as it bounds ratio-nality to the world rather than treats the two as fundamentally separate (Todd & Gigerenzer, 2003). Thus, the ecological approach differs significantly from both clas-sical rationality and the heuristics-and-biases tradition, and it offers what we consider another way of thinking about rationality:

There is a third way to look at inference, focusing on the psychological and ecolog-ical rather than on logic and probability theory. This view questions classical ratio-nality as a universal norm and thereby questions the very definition of good rea-soning on which both the Enlightenment and heuristics-and-biases views were built. (Gigerenzer S Goldstein, 1996, p. 651)

But in what sense does ecological ratio-nality differ from classical rationality? The apparent strength of rational principles of inference stem from their generality and their formal justification by way of prob-ability theory, and ultimately information theory (Li & Vitnyi, 1997). For example, Occam's razor tells us to prefer simpler explanations over complex ones, and this principle proves productive when fitting a polynomial to a set of data points, choos-ing between scientific theories, or describ-ing the behavior of human visual percep-tion (Sober, 1975). That such principles hold across diverse tasks that are seemingly unre-lated in their formulation, their physical characteristics, and importantly, the envi-ronmental context, is seen as evidence of their strength.

However, if our concern is the study of human and animal cognition, then there are good reasons to view these apparent

strengths as weaknesses. Neither humans nor animals uniformly adhere to overarch-ing principles because, as we will argue, such principles make cognitive demands that can-not always be met. Furthermore, the gener-ality of rational principles is in large part due to abstractions away from specific aspects of the task. Sometimes specific considerations, such as response time, are significant and may outweigh more general considerations such as predictive accuracy. We suspect that no single and universal measurement can characterize what is functional for an organ-ism for all tasks, and in this sense we should be weary of proposals that collapse the prob-lem characterizing rational choice down to a single yardstick. The concept of ecological rationality accepts the deep problems that arise with universal and abstract principles of rationality and works with them.

First of all, an important distinction to consider is that rational principles of infer-ence are normative vehicles for judging inductive inferences. As such, they are inert with respect to how, in processing terms, organisms should arrive at inferences. Her-bert Simon (1990, chap. 2) made the dis-tinction between substantive and procedural rationality. Although some formulations of rational behavior consider procedural ratio-nality, rational theories of inductive infer-ence are typically claims about substantive rationality. Substantive rational principles are yardsticks for judging inferences; they do not tell us, in mechanistic terms, how to arrive at inductive inferences. It is use-ful, therefore, to distinguish between the substantive problem of induction and the procedural problem of induction. This dis-tinction is essential when we come to con-sider fundamental computational limits that make realizing rational norms in process-ing terms often intractable, if not provably uncomputable. Note that this problem -the dichotomy between what is rational and what is computationally achievable -extends beyond the particular details of the cognitive system. It is an issue for all com-putational processes carrying out inductive inferences. We mention the processing issue here because, if what is deemed rational is

SITUATING RATIONALITY 333

fundamentally unobtainable by organisms, then perhaps our motives for adopting such a concept of rationality should be ques-tioned.

Another distinction we will draw con-trasts the substantive problem of induc-tion and the cognitive-ecological problem of induction. This distinction will lead us to consider how behaving rationally in the classical sense (the substantive problem) and behaving adaptively (a species-specific cognitive-ecological problem) are not neces-sarily the same endeavor. For instance, the inference suggested by the most probable prediction (and therefore the most rational one) may require expending considerable time, memory, and processing resources compared with an inference arrived at quickly and on the basis of a cognitively undemanding decision process requiring minimal information; moreover, the latter may be only marginally inferior in proba-bilistic terms. Adaptive behavior, that which fits the functional requirements of an organ-ism, is unlikely to be characterized using probability theory and logic given that "peo-ple satisfice - look for good enough solu-tions - instead of hopelessly searching for the best" (Simon, 1990, p. 17). Put differ-ently, the payoff function with which we measure the efficacy of a decision maker is not simply how accurate inferences are; it also must consider ecological factors that take into account the structure of the task, the criticality of response time, and factors contributing to cognitive effort. Now, in this section the question we ask is, Does one rationality fit all tasks?

4.1. Living Up to Expectation: From Substantive to Procedural Rationality

Substantive rationality refers to rational prin-ciples such as Bayesian inference or Occam's razor. Procedural rationality, in contrast, considers what can be accomplished when one takes into account the processing steps required to arrive at an inference. Once an inference has been made, the role of sub-stantive rationality is clear: it provides a cri-terion with which to judge this inference

against others. By considering the problem of procedural rationality, we are taking one step toward situating the task of human decision making. Constraints on realization can transform the nature of the problem, or, as Herbert Simon (1990) put it: "at each step toward realism, the problem gradu-ally changes from choosing the right course of action (substantive rationality) to find-ing a way of calculating, very approximately, where a good choice of action lies (procedu-ral rationality)" (pp. 26-27). In stressing this difference between substantive and proce-dural rationality in this way, we do not want give the impression that substantive theo-ries of rationality are wrong. The issue is how appropriately a given notion of rational-ity fits the question at hand. Simon's point, and the point we are developing here, is that, for the task of understanding cogni-tion, procedural rationality is more appro-priate than substantive rationality. Again, this does not imply that substantive theo-ries of rationality, such as Bayesianism, are the wrong tool for studying cognition, but rather that such principles are inherently limited because they neglect the fact that organisms are constrained in their ability to process information.

Some tasks, because of their inherent dif-ficulty, lack a clear notion of substantive rationality. For example, choosing which of two candidate moves in chess is most likely to lead to a win can, in the general case, be only an estimate. Unsurprisingly, this inher-ent difficulty makes the procedural prob-lem of playing chess even harder. In other words, if we lack a clear formulation of ratio-nal choice in the first place, then we can-not expect that procedural solutions con-form to this rational expectation. Although chess is a supremely unenlightening exam-ple of human inference that, at best, may tell us about the fringes of human cogni-tion (Chomsky, 2000, p. 161), there are many other common human decisions that are similarly difficult - or impossible - to ratio-nally analyze fully, such as deciding which of two potential mates to woo (Gigerenzer & Todd, 1999). Furthermore, there are tasks for which we do have a clear notion of

334 HENRY BRIGHTON A N D PETER M. T O D D

substantive rationality, yet procedural ratio-nality still forces us to rethink what is achiev-able from the perspective of the organism. This is the issue we will focus on.

Consider the modern formulation of Occam's razor: hypotheses that recode observations in such a way to minimize encoding length should be chosen over oth-ers (Grnwald, 2005: Hutter, 2005; Li & Vitnyi, 1997; Rissanen, 1989). Such a for-mulation has been proposed as a unifying principle of the cognitive system (Chater & Vitnyi, 2003; see also Feldman, 2003) and "suggests a possible (although of course par-tial) account of the remarkable success of the cognitive system in prediction, under-standing, and acting in an uncertain and complex environment: that cognitive pro-cesses search for simplicity" (Chater, 1999, p. 283). Vet paradoxically, adherence to simplicity principles can be so complex in processing terms that it "will not be possible in general" (Chater, 1999, p. 283). As another example of the substantive-procedural gap, in both perception and action it is pro-posed that a "striking observation.. . is the myriad ways in which human observers act as optimal Bayesian observers" (Knill &. Pouget, 2004, p. 712). Once again, however, it is acknowledged that, "owing to the com-plexity of the task, unconstrained Bayesian inference is not a viable solution for com-putation in the brain" (Knill & Pouget, 2004, p. 718). Here we see the struggle between the desire to find concise universal laws of rational behavior and the processing diffi-culties that accompany these principles. To see why this problem occurs, and how the concept of ecological rationality can help to clarify issues, we will consider the cog-nitive task of categorization. Categorization requires making inductive inferences, it is a ubiquitous ability of humans and other ani-mals, and it offers a useful example for illus-trating the dichotomy between procedural and substantive rationality.

For the categorization task, Bayesian inference and modern castings of Occam's razor provide well-defined and precise rational criteria with which to compare one potential category judgment against

another. Given a series of observations, these criteria tell us, for any two candidate category explanations, which is the "best" category description. Stepping back from any specific solutions that the cognitive sys-tem may employ, it will prove useful to con-sider first the more general class of computa-tional procedures for performing inductive inferences. Machine learning is the study of procedural solutions to inference prob-lems (Mitchell, 1997), and therefore offers a useful source of insight into the rela-tionship between substantive and procedu-ral rationality (which is sometimes termed approximate rationality within artificial intel-ligence; see Russell & Norvig, 1995). Like economic and psychological research, artifi-cial intelligence and, in particular, machine learning often appeal to classical notions of rational behavior (Goodie, Ortmann, Davis, Bullock, &. Werner, 1999). For example, using algorithmic information theory to fix a universal (i.e., problem independent, parameterless, and rationally motivated by way of Occam's razor) prior probabil-ity distribution on the hypothesis space (Solomonoff, 1964), Bayesian-inspired mod-els of inductive reasoning have led some to explore "a theory for rational agents act-ing optimally in any environment" (Hutter, 2005, p. 24). Unfortunately, when realized in procedural terms, such theories partly assume the "availability of unlimited com-putational resources" (Hutter, 2005, p. v), Even with extremely liberal constraints on resources, universally applicable laws of rationality, in processing terms, are difficult or even impossible to achieve.

In practice, machine-learning research places more conservative bounds on what is considered practical by often narrowing its interests to the class of computation-ally tractable algorithms. Computationally tractable algorithms are those that place computational demands on time and stor-age space resources that grow as polynomial function of the size of the problem (e.g., sorting a series of n numbers into ascend-ing order is deemed tractable because it can be achieved by an algorithm requiring time and space polynomial in w). Under these

SITUATING RATIONALITY 335

constraints, a widespread realization, and arguably an axiom within machine learn-ing, is that procedural adherence to rational expectation breaks down. One useful con-cept in thinking about ad hoc adherence to the rational ideal is inductive bias, which refers to any basis on which one explana-tion is chosen over another, beyond simple adherence to the observations (Gordon & Desjardins, 1995; Mitchell, 1997).' An appro-priately formalized Occam's-razor principle is a bias, for example, but typically bias occurs as a result of restrictions on the rep-resentational power of the hypothesis space (the set of explanations the algorithm can consider) and bow thoroughly the hypoth-esis space can be searched (the manner in which the search for the optimal hypothesis is approximated).

The important point here is that the inductive bias of an algorithm typically reflects concrete issues of realization such as the particular characteristics and assump-tions imposed by the implementation of the algorithm (e.g., the nature of hypothe-sis space) rather than a rationally motivated normative bias, such as Occam's razor. Con-sequently, and as a result of tractability and computability constraints, the procedural problem of induction has a different charac-ter from that suggested by the rational prob-lem. The combination of considering non-trivial tasks in conjunction with tractability constraints limits researchers to a vaguer adherence to rational expectations, where algorithms "should process the finite sample to obtain a hypothesis with good general-ization ability under a reasonably large set of circumstances" (Kearns, 1999, p. 159). For machine learning, this breakdown in adher-ence to the rational, or optimal, outcome is simply a well-established fact. Algorithmic solutions are, to varying degrees, focused rather than general, and their performance is adequate rather than normative:

Induction is not unbridled or uncon-strained. Indeed, decades of work in machine learning makes abundantly clear that there is no such thing as a general pur-pose learning algorithm that works equally

well across domains. Induction may be the name of the game, but constraints are the rules that we play by. (Elman, 2003, from ms.)

Because uniform adherence to a norma-tive criterion is taken as a goal, alleviating the discrepancy between this goal and what can be achieved in practice can partially be overcome through the selective choice of learning algorithms, dependent on the task structure. For example, given knowl-edge of the tasks for which some learning algorithm A performs better any other algo-rithm, then A could be chosen over all oth-ers when such a task is encountered. This problem is known as the selective superior-ity problem. It remains largely unexplored, and only limited practical progress has been made in addressing it (e.g., Kalousis, Gama, & Hilario, 2004). Unfortunately, for anything other than trivial tasks, machine learning tells us that there is a gap between sub-stantive and procedural rationality. This gap exists because tractable algorithms tend to focus their performance on some problems at the expense of others. Crucially, most tasks are not trivial in the sense we mean here and require the expenditure of con-siderable computational effort to find good, let alone optimal, solutions. For example, the apparently trivial task of recognizing a face is extremely difficult from a computa-tional standpoint: a full integration of the many facial properties that need to be con-sidered is computationally intractable. In the study of artificial intelligence, which seeks engineering solutions for tasks like these, optimal solutions are almost always unob-tainable (for discussion, see Reddy, 1988). Likewise, the study of cognition rarely reduces to the study of computationally triv-ial problems.

This brief discussion of machine learning provides us with some conceptual tools with which to consider the cognitive problem. The cognitive problem of induction is a par-ticular instance of the procedural problem of achieving rationality, and one that is sub-ject to hard biological/cognitive constraints rather than the more abstract notion of

HENRY BRIGHTON A N D PETER M. T O D D 3 3 6

computational tractability encountered pre-viously. As for machine learning's algorith-mic rationality problem, the general human cognitive problem is also broken down into tasks with particular characteristics. For instance, tasks spanning low-level aspects of the visual system and high-level tasks such as concept learning and categorization are commonly viewed as inductive tasks (e.g., Chater & Vitnyi, 2003; Sober, 1975]. How-ever, unlike a general constraint such as computational tractability, these cognitive tasks are likely to work within very different and more stringent constraints, such as the physical limitations of the underlying bio-logical machinery; constraints imposed by other cognitive systems also using resources in the mind; and limits on attention, working memory, and the like. Thus, different cog-nitive tasks are likely to yield to processing solutions of different forms.

4.2. Situating Decision Making: Confronting the Cognitive-Ecological Problem

Processing makes demands on computa-tional resources. Placing constraints on these resources limits the degree to which ratio-nal expectations can be met, and in partic-ular, processes tend to be focused on some instances of the task with respect to their ability to adhere to the demands set by sub-stantive rationality. Specific kinds of con-straint impose specific kinds of focus and, as Simon argued (1990, 1996), a full consid-eration of cognitive limitations leads us to consider boundedly rational processes. On the one hand, constraints beyond those of an abstract computational nature would appear to limit even further the degree to which rational expectations are likely to be met, and, as a result, the human cognitive system needs to pull a neat trick if it is to measure up to the demands of rationality under these terms. On the other hand, we will argue that there are good reasons to view cognitive limitations not as barriers bu t as enablers of robust inference. Limitations can be viewed as adaptive constraints in th