The target article urges that our criteria of rationality shouldn't ignore resource limitations, indeed that, properly understood, the demands of effectively deploying limited computational resources provides a unifying basis for recent work on how humans and other animals deviate from traditional models of rationality. The argument is worryingly silent on two related problems – processing utility representations, and making action out of a body.
Traditional rational actor models, and the refinements discussed in the target article (Equations 1 to 4) include a utility function. Their appearance in accounts of rationality isn't surprising, because instrumentally understood rationality is a matter of effective pursuit of some goals. Influential arguments defend the view that an effective agent will, among other things, have goals that satisfy certain requirements of consistency (Ramsay Reference Ramsay1931), and behave as if she assigned – and updated – subjective probabilities to current and future states of the world (including states consequent on her own actions), and selected actions that maximised expected utility (Savage Reference Savage1954). Okasha places these lines of thinking in an evolutionary context, to argue that an agent that acts and chooses as if performing Bayesian updating is an optimal agent, and hence a plausible target, at least sometimes, for natural selection (Okasha Reference Okasha2013). These arguments are typically understood behaviourally. Unlike the target article they defend claims about what effective agents do, not how they work.
Clearly enough, though, one straightforward way to act as if having beliefs updated in certain ways, and a utility function with certain properties, is to actually have those beliefs and preferences. More specifically, and leaving representations of the world to one side, whether a cognitive architecture is rational or resource-rational, if it is going to try to maximise utility, it has to build and maintain representations that convert and integrate the different dimensions of cost and return that matter to the agent into a consistent measure of utility, or common currency.
This has implications for cognitive architecture: If a “biologically feasible mind” (which the target article explicitly aims at) is to perform operations involving utility, then it has, somehow, to generate states that represent the expected returns from actions or strategies, future world states, life history segments, etc. These complications are sometimes ignored in models of rationality, which fix utility by fiat (“consider an agent valuing one slice of pizza as much as two ice-cream cones”) in order to focus – like the target article – on other technicalities. Most living agents, though, have to deal with a heterogeneous mixture of costs and returns. The costs include direct expenditure of energy, the depletion of specific “fuels” or resources such as water and salt, as well as time and exposure to various risks. The returns include hydration, nutrition, rest, access to mating opportunities, acquisition of nesting materials or control of a nesting site, and so forth. A plausible list of only the primary reinforcers in humans (Rolls Reference Rolls2013) enumerates almost 50, some of them – such as “hormones” and “facial expressions” – with many variations falling into them. Costs and returns in these reinforcers are clumped together in heterogeneous bundles out in the world. Even in ideal circumstances, where important facts can be cheaply and reliably detected, converting the many dimensions of cost and return into utility values is likely to be difficult and resource-hungry. Circumstances aren't generally ideal, in part because much of the living world consists of rivals and competitors rather than allies. In both appearance and behaviour plants and animals often take steps to conceal or misrepresent their identity and likely behaviour. Sterelny (Reference Sterelny2003) called this informational “hostility,” and it means that merely tracking costs and returns will sometimes be subject to trade-offs between cost and accuracy involving a further kind of resource limitation.
Using utility to select actions also requires sensitivity to the demands of controlling the body. These complications are ignored in many models of rationality, which take functional actions as primitive. The target article itself doesn't mention the body, and it does, perhaps revealingly, describe the resource rational “brain B interacting with the environment” (caption to Figure 1). This isn't how it works at all. In fact the stock of actions of a big complex animal like a human depends on a large structured array of muscles and other effectors, which only produces functional activity when subject to appropriate and sometimes complicated patterns of activation and suppression. The components of the actions themselves have their own metabolic and opportunity costs. That is to say, the cognitive demands of action production (getting the array of capacities to do this rather than that, or to do anything functional at all) aren't independent of the problem of trying to maximise utility (Spurrett Reference Spurrett2019).
Both of these problems can vary in their demands. The overheads of constructing utility representations increase with the number of types of rewards and costs to which the system is to be responsive, as well as the costs of detecting cues of them with acceptable accuracy. (Detecting an acceptable egg to brood can cost more if your species is exploited by cuckoos.) The overheads of constructing the actions that a body is mechanically capable of increase with the number of degrees of freedom available, and how many need to be coordinated to produce functional activity. Attempting to maximise utility in an agent with a real body, and a real suite of sensory transducers (external and internal) requires facing up to these trade-offs.
There's a choice to be made here: Either admit to developing an account of rationality tailored to agents with a telekinetic action repertoire and roughly magical power to detect predictors of utility, so that the brain really does interact with the environment, or take the body seriously. The body, as a source of various channels of information about the world, the thing whose needs are the dimensions out of which utility is made, and as the thing that has to be controlled to produce action, is home to additional important kinds of resource-constraints.
You have
Access
The target article urges that our criteria of rationality shouldn't ignore resource limitations, indeed that, properly understood, the demands of effectively deploying limited computational resources provides a unifying basis for recent work on how humans and other animals deviate from traditional models of rationality. The argument is worryingly silent on two related problems – processing utility representations, and making action out of a body.
Traditional rational actor models, and the refinements discussed in the target article (Equations 1 to 4) include a utility function. Their appearance in accounts of rationality isn't surprising, because instrumentally understood rationality is a matter of effective pursuit of some goals. Influential arguments defend the view that an effective agent will, among other things, have goals that satisfy certain requirements of consistency (Ramsay Reference Ramsay1931), and behave as if she assigned – and updated – subjective probabilities to current and future states of the world (including states consequent on her own actions), and selected actions that maximised expected utility (Savage Reference Savage1954). Okasha places these lines of thinking in an evolutionary context, to argue that an agent that acts and chooses as if performing Bayesian updating is an optimal agent, and hence a plausible target, at least sometimes, for natural selection (Okasha Reference Okasha2013). These arguments are typically understood behaviourally. Unlike the target article they defend claims about what effective agents do, not how they work.
Clearly enough, though, one straightforward way to act as if having beliefs updated in certain ways, and a utility function with certain properties, is to actually have those beliefs and preferences. More specifically, and leaving representations of the world to one side, whether a cognitive architecture is rational or resource-rational, if it is going to try to maximise utility, it has to build and maintain representations that convert and integrate the different dimensions of cost and return that matter to the agent into a consistent measure of utility, or common currency.
This has implications for cognitive architecture: If a “biologically feasible mind” (which the target article explicitly aims at) is to perform operations involving utility, then it has, somehow, to generate states that represent the expected returns from actions or strategies, future world states, life history segments, etc. These complications are sometimes ignored in models of rationality, which fix utility by fiat (“consider an agent valuing one slice of pizza as much as two ice-cream cones”) in order to focus – like the target article – on other technicalities. Most living agents, though, have to deal with a heterogeneous mixture of costs and returns. The costs include direct expenditure of energy, the depletion of specific “fuels” or resources such as water and salt, as well as time and exposure to various risks. The returns include hydration, nutrition, rest, access to mating opportunities, acquisition of nesting materials or control of a nesting site, and so forth. A plausible list of only the primary reinforcers in humans (Rolls Reference Rolls2013) enumerates almost 50, some of them – such as “hormones” and “facial expressions” – with many variations falling into them. Costs and returns in these reinforcers are clumped together in heterogeneous bundles out in the world. Even in ideal circumstances, where important facts can be cheaply and reliably detected, converting the many dimensions of cost and return into utility values is likely to be difficult and resource-hungry. Circumstances aren't generally ideal, in part because much of the living world consists of rivals and competitors rather than allies. In both appearance and behaviour plants and animals often take steps to conceal or misrepresent their identity and likely behaviour. Sterelny (Reference Sterelny2003) called this informational “hostility,” and it means that merely tracking costs and returns will sometimes be subject to trade-offs between cost and accuracy involving a further kind of resource limitation.
Using utility to select actions also requires sensitivity to the demands of controlling the body. These complications are ignored in many models of rationality, which take functional actions as primitive. The target article itself doesn't mention the body, and it does, perhaps revealingly, describe the resource rational “brain B interacting with the environment” (caption to Figure 1). This isn't how it works at all. In fact the stock of actions of a big complex animal like a human depends on a large structured array of muscles and other effectors, which only produces functional activity when subject to appropriate and sometimes complicated patterns of activation and suppression. The components of the actions themselves have their own metabolic and opportunity costs. That is to say, the cognitive demands of action production (getting the array of capacities to do this rather than that, or to do anything functional at all) aren't independent of the problem of trying to maximise utility (Spurrett Reference Spurrett2019).
Both of these problems can vary in their demands. The overheads of constructing utility representations increase with the number of types of rewards and costs to which the system is to be responsive, as well as the costs of detecting cues of them with acceptable accuracy. (Detecting an acceptable egg to brood can cost more if your species is exploited by cuckoos.) The overheads of constructing the actions that a body is mechanically capable of increase with the number of degrees of freedom available, and how many need to be coordinated to produce functional activity. Attempting to maximise utility in an agent with a real body, and a real suite of sensory transducers (external and internal) requires facing up to these trade-offs.
There's a choice to be made here: Either admit to developing an account of rationality tailored to agents with a telekinetic action repertoire and roughly magical power to detect predictors of utility, so that the brain really does interact with the environment, or take the body seriously. The body, as a source of various channels of information about the world, the thing whose needs are the dimensions out of which utility is made, and as the thing that has to be controlled to produce action, is home to additional important kinds of resource-constraints.