Leider and Griffiths’ (L&G) project, to offer an explicit framework for relativizing assessments of rationality simultaneously to cognitive processing constraints and environmental affordances, represents important progress. It significantly clarifies the basis for unification between mechanism-driven and solution-driven disciplines and methodologies, as they say. But, as the framework is extended and refined, two outstanding issues merit consideration: (1) human brains do not merely adapt to their environments, but co-process information with their environments, particularly with its social aspects; and (2) L&G's idealization of disciplines as standing in a hierarchy of abstraction from mechanism details is a somewhat misleading simplification of methodological reality.
L&G's core Equation 4 takes the environment (E) as a fixed constraint on optimal heuristic selection. This is reasonable in light of the long time-scale for learning that their discussion indicates that they have in mind, reflected in their comment that evolution and cognitive development “solve the constrained optimization problem defined in Equation 3” (sect. 3, para. 5). The framework obviously allows for environmental variation, across time or space, to be modeled and analyzed using comparative statics. Furthermore, their inclusion of the information term I on the left-hand side of Equation 4 recognizes that learning encoded in the genome is refined by learning in the phenome. However, the model seems to presuppose that cognitive processing is all done in the brain, because there is no interaction term involving all of h, E, and B (heuristics, environment, and brain).
This may be a reasonable idealization where most cognitive systems are concerned. But, it might be seriously misleading in the case of humans equipped with writing, art, and mathematics, who have populated their environments with technologies that actively process information in conjunction with inboard cognition. Obvious examples include external computing devices, but these are not the main source of potential deep complication for L&G's model. Though, the relationship between a person and a machine she uses may be dynamically interactive, in non-exotic cases the extent of such dynamical coupling is both limited and specifiable; and, as noted above, this is all that is required for analyzing variation by means of comparative statics. The more serious challenge arises from the abstract technology of social institutions. Ecologically, humans are arguably most strongly distinguished from other highly intelligent animals by their use of shared information-processing routines that are encoded in rules, norms, and institutionalized procedures, which individuals exploit by mixtures of faithful and noisy compliance and deliberate modification. It is not immediately evident how L&G's Equation 4 should be modified to parcel elements of such social heuristic (or optimizing) processes into inboard and outboard elements. Put in terms of a simple example: if an entrepreneur generally follows her venture capitalist's boilerplate advice, but distorts it through a mix of subjective probability weighting on risks and explicit private knowledge, and this in turn influences all participants’ models of their market, which elements are to be included in B and which in E?
L&G quote Simon's salutary observation that “the environment may lie, in part, within the skin of the biological organism” (sect. 3, para. 7). Their model reflects this insight. But, then a complementary point, as emphasized by theorists such as Clark (Reference Clark1997) and Sterelny (Reference Sterelny2003; Reference Sterelny2012), is that some cognitive processing occurs outside of the skin. I stress again that I am not referring here to information processing that is largely exogenous to psychological mechanisms, as in the computation of graphical representations by a statistics package or of prices by a market. To call such processing “cognitive” would involve capture by a metaphor. The point, rather, is that as Clark (Reference Clark2003) emphasizes, people expand their intelligence by coupling their brains with the representational and active computational tools that they collectively operate; and the boundary between individual and social resource-rationality is, in the context of conserved engineering achievements, blurred to the extent of collapse. Reliance on outboard processing for much of the very sophisticated information processing characteristic of humans is plausibly an essential requirement arising from metabolic constraints.
This concern is related to an assumption L&G make explicit later in their paper about the relationships among disciplines. They describe economics (along with AI) as simplifying and idealizing models of the mind. Behavioral economists who seek strong unification between their discipline and psychology are likely to be comfortable with this, but it obscures important methodological differences. Psychometrics is, to a first approximation, the statistical theory of measures of construct validity because psychologists aim to infer ‘hidden’ mechanisms from observations of behavior, and thus need to exclude ‘confounding’ elements of E when experimentally focused on B. By contrast, experimental economists tend to deliberately undermine the importance of the E/B distinction by adding new treatments where psychologists would seek to block out a “confound.” This explains why the econometrics of the lab is essentially the statistical theory of structural model identification and estimation. Although, economists are of course students of information-processing, it does not seem apt to depict them as studying abstract, idealized minds. This might look like merely philosophical quibbling. But, in fact it anchors the concern about the clarity of the E/B distinction in terms of practical modeling: as an economist I would need detailed reassurance that if standard, working structural models are to be constrained by Equation 4, this would not require solving identification problems that economists have worked hard to bypass.
I close with an example. Cumulative prospect theory, which is certainly an idealized model of mind just as L&G say it is, is an awkward tool for the economist's lab because it locks in parameters that are extremely difficult to identify (Harrison and Swarthout Reference Harrison and Swarthout2016), and which lack principled theoretical generalization. And almost all of the relevant empirical estimation work can instead be done using a rank-dependent utility specification that allows for subjective decision weights that fail to track objective probabilities. Do the subjective weightings in question come from B or from (the social) E (see Harrison and Ross Reference Harrison and Ross2017)? Does the economist really need to care about the answer?
You have
Access
Leider and Griffiths’ (L&G) project, to offer an explicit framework for relativizing assessments of rationality simultaneously to cognitive processing constraints and environmental affordances, represents important progress. It significantly clarifies the basis for unification between mechanism-driven and solution-driven disciplines and methodologies, as they say. But, as the framework is extended and refined, two outstanding issues merit consideration: (1) human brains do not merely adapt to their environments, but co-process information with their environments, particularly with its social aspects; and (2) L&G's idealization of disciplines as standing in a hierarchy of abstraction from mechanism details is a somewhat misleading simplification of methodological reality.
L&G's core Equation 4 takes the environment (E) as a fixed constraint on optimal heuristic selection. This is reasonable in light of the long time-scale for learning that their discussion indicates that they have in mind, reflected in their comment that evolution and cognitive development “solve the constrained optimization problem defined in Equation 3” (sect. 3, para. 5). The framework obviously allows for environmental variation, across time or space, to be modeled and analyzed using comparative statics. Furthermore, their inclusion of the information term I on the left-hand side of Equation 4 recognizes that learning encoded in the genome is refined by learning in the phenome. However, the model seems to presuppose that cognitive processing is all done in the brain, because there is no interaction term involving all of h, E, and B (heuristics, environment, and brain).
This may be a reasonable idealization where most cognitive systems are concerned. But, it might be seriously misleading in the case of humans equipped with writing, art, and mathematics, who have populated their environments with technologies that actively process information in conjunction with inboard cognition. Obvious examples include external computing devices, but these are not the main source of potential deep complication for L&G's model. Though, the relationship between a person and a machine she uses may be dynamically interactive, in non-exotic cases the extent of such dynamical coupling is both limited and specifiable; and, as noted above, this is all that is required for analyzing variation by means of comparative statics. The more serious challenge arises from the abstract technology of social institutions. Ecologically, humans are arguably most strongly distinguished from other highly intelligent animals by their use of shared information-processing routines that are encoded in rules, norms, and institutionalized procedures, which individuals exploit by mixtures of faithful and noisy compliance and deliberate modification. It is not immediately evident how L&G's Equation 4 should be modified to parcel elements of such social heuristic (or optimizing) processes into inboard and outboard elements. Put in terms of a simple example: if an entrepreneur generally follows her venture capitalist's boilerplate advice, but distorts it through a mix of subjective probability weighting on risks and explicit private knowledge, and this in turn influences all participants’ models of their market, which elements are to be included in B and which in E?
L&G quote Simon's salutary observation that “the environment may lie, in part, within the skin of the biological organism” (sect. 3, para. 7). Their model reflects this insight. But, then a complementary point, as emphasized by theorists such as Clark (Reference Clark1997) and Sterelny (Reference Sterelny2003; Reference Sterelny2012), is that some cognitive processing occurs outside of the skin. I stress again that I am not referring here to information processing that is largely exogenous to psychological mechanisms, as in the computation of graphical representations by a statistics package or of prices by a market. To call such processing “cognitive” would involve capture by a metaphor. The point, rather, is that as Clark (Reference Clark2003) emphasizes, people expand their intelligence by coupling their brains with the representational and active computational tools that they collectively operate; and the boundary between individual and social resource-rationality is, in the context of conserved engineering achievements, blurred to the extent of collapse. Reliance on outboard processing for much of the very sophisticated information processing characteristic of humans is plausibly an essential requirement arising from metabolic constraints.
This concern is related to an assumption L&G make explicit later in their paper about the relationships among disciplines. They describe economics (along with AI) as simplifying and idealizing models of the mind. Behavioral economists who seek strong unification between their discipline and psychology are likely to be comfortable with this, but it obscures important methodological differences. Psychometrics is, to a first approximation, the statistical theory of measures of construct validity because psychologists aim to infer ‘hidden’ mechanisms from observations of behavior, and thus need to exclude ‘confounding’ elements of E when experimentally focused on B. By contrast, experimental economists tend to deliberately undermine the importance of the E/B distinction by adding new treatments where psychologists would seek to block out a “confound.” This explains why the econometrics of the lab is essentially the statistical theory of structural model identification and estimation. Although, economists are of course students of information-processing, it does not seem apt to depict them as studying abstract, idealized minds. This might look like merely philosophical quibbling. But, in fact it anchors the concern about the clarity of the E/B distinction in terms of practical modeling: as an economist I would need detailed reassurance that if standard, working structural models are to be constrained by Equation 4, this would not require solving identification problems that economists have worked hard to bypass.
I close with an example. Cumulative prospect theory, which is certainly an idealized model of mind just as L&G say it is, is an awkward tool for the economist's lab because it locks in parameters that are extremely difficult to identify (Harrison and Swarthout Reference Harrison and Swarthout2016), and which lack principled theoretical generalization. And almost all of the relevant empirical estimation work can instead be done using a rank-dependent utility specification that allows for subjective decision weights that fail to track objective probabilities. Do the subjective weightings in question come from B or from (the social) E (see Harrison and Ross Reference Harrison and Ross2017)? Does the economist really need to care about the answer?