Lieder and Griffiths argue that human cognition can be understood using much of the same framework that we would use to understand how a cash register works: in terms of the system's function and its resources (Marr Reference Marr1982). But because people are different from cash registers, perhaps the best framework to understand one is not all that useful to understand the other. A major problem stems from the different ways in which these systems solve the correspondence problem – the problem of generating representations of the environment that correspond to the actual environment (Hammond Reference Hammond, Conolly, Hammond and Arkes2000).
The cash register does not need to solve this problem – it has already been solved by its programmer. Because the programmer decided what the function of the machine will be, she presumably included all the necessary algorithms that can generate representations of the desired aspects of the world. For example, the machine needs an algorithm that generates representations of the value of items based on the key-presses of its user. Although such a heuristic will allow the cash register to represent this aspect of the environment well, it will also only ever represent these aspects.
Lieder and Griffiths suggest that evolution acted in a similar way to this programmer, and endowed people with algorithms that are optimised to generate representations of the environment. Thus, people do not need to solve the correspondence problem, because it has already been solved by evolution. According to this view, the difference between cash registers and people is quantitative: people have more algorithms – an “adaptive toolbox” of heuristics that generate environmental representations for them.
The shared underlying idea is that representation generation effectively consists of the matching of environmental input with already stored representations of potential environments (cf., Lieder & Griffiths Reference Lieder and Griffiths2017). For the cash register, this amounts to, for example, the deterministic matching of the key-press with the representation of the corresponding number. For people, the inputs are environmental features that were deemed relevant by evolution, some summary of which are then matched to the representation of the most similar problem. The main aim of resource-rational analysis is to find the solution to this problem with the highest expected value while also taking computational costs into account. This approach presupposes a clearly definable function for the system (i.e., a fixed environmental representation), which is defined by the researcher (target article, Box 2, Step 1).
But what if human cognition is not just a bag of tricks? It is not difficult to conceive how evolution would have favoured a cognitive system that can represent the environment more flexibly. Humans seem to be capable of representing any possible environment (and even impossible ones) and use that knowledge to learn and make decisions – in other words, representation generation in humans seems to be universal (cf., Deutsch Reference Deutsch2011). Such a view of human cognition suggests that people are qualitatively different from the cash register in the sense that they can actually try to solve the correspondence problem. In fact, it suggests that people are more similar to scientists than to cash registers; and just like scientists, people need to generate and select the best out of many possible and plausible representations.
The flexibility in how people generate representations is demonstrated both by observations that their representations of the same environment can show considerable differences (e.g., Gaissmaier & Schooler Reference Gaissmaier and Schooler2008; Schulze & Newell Reference Schulze and Newell2016), and by observations that such representations can be improved on (e.g., Szollosi et al. Reference Szollosi, Liang, Konstantinidis, Donkin and Newell2019). The potential explanations that (a) these are in fact not new or improved representations, but result from the misapplication of strategies that evolved in a different evolutionary milieu; or (b) that people assess all potentially relevant features of the environment are both unsatisfactory. The former increases the flexibility of the model to an extent that there is almost nothing it cannot account for. The latter is computationally impossible.
This is not to say that people do not rely on heuristics to generate representations of the environment. Our argument is only that we should appreciate the flexibility of this generation process in our models, instead of substituting it with fixed representations based on flexible assumptions of the researcher. A promising way for such investigations would be the use of simple yet diagnostic manipulations of purportedly relevant features of the environment followed by thorough probing of people's knowledge about these features (e.g., Tran et al. Reference Tran, Vul and Pashler2017; also see, Newell & Shanks Reference Newell and Shanks2014).
The usefulness of resource-rational analysis hinges on the assumed similarity in representation generation between cash registers and people, because only under such conditions can the purpose of the system be clearly defined. We argued against this assumption: people seem capable of generating representations of anything, whereas cash registers can only represent things that they were programmed to represent. This difference leads to a radically different computational level question about cognition. Instead of asking what the purpose of people's cognitive mechanisms are in terms of prototypical evolutionary or learning environments, we could ask what purpose they serve in achieving the universality of representation generation. Such universality makes characterising the function of human cognition elusive, because with every new representation of a problem, the person generates a new function. Theories of human cognition need to clarify the development of purpose not merely presume its existence.
Lieder and Griffiths argue that human cognition can be understood using much of the same framework that we would use to understand how a cash register works: in terms of the system's function and its resources (Marr Reference Marr1982). But because people are different from cash registers, perhaps the best framework to understand one is not all that useful to understand the other. A major problem stems from the different ways in which these systems solve the correspondence problem – the problem of generating representations of the environment that correspond to the actual environment (Hammond Reference Hammond, Conolly, Hammond and Arkes2000).
The cash register does not need to solve this problem – it has already been solved by its programmer. Because the programmer decided what the function of the machine will be, she presumably included all the necessary algorithms that can generate representations of the desired aspects of the world. For example, the machine needs an algorithm that generates representations of the value of items based on the key-presses of its user. Although such a heuristic will allow the cash register to represent this aspect of the environment well, it will also only ever represent these aspects.
Lieder and Griffiths suggest that evolution acted in a similar way to this programmer, and endowed people with algorithms that are optimised to generate representations of the environment. Thus, people do not need to solve the correspondence problem, because it has already been solved by evolution. According to this view, the difference between cash registers and people is quantitative: people have more algorithms – an “adaptive toolbox” of heuristics that generate environmental representations for them.
The shared underlying idea is that representation generation effectively consists of the matching of environmental input with already stored representations of potential environments (cf., Lieder & Griffiths Reference Lieder and Griffiths2017). For the cash register, this amounts to, for example, the deterministic matching of the key-press with the representation of the corresponding number. For people, the inputs are environmental features that were deemed relevant by evolution, some summary of which are then matched to the representation of the most similar problem. The main aim of resource-rational analysis is to find the solution to this problem with the highest expected value while also taking computational costs into account. This approach presupposes a clearly definable function for the system (i.e., a fixed environmental representation), which is defined by the researcher (target article, Box 2, Step 1).
But what if human cognition is not just a bag of tricks? It is not difficult to conceive how evolution would have favoured a cognitive system that can represent the environment more flexibly. Humans seem to be capable of representing any possible environment (and even impossible ones) and use that knowledge to learn and make decisions – in other words, representation generation in humans seems to be universal (cf., Deutsch Reference Deutsch2011). Such a view of human cognition suggests that people are qualitatively different from the cash register in the sense that they can actually try to solve the correspondence problem. In fact, it suggests that people are more similar to scientists than to cash registers; and just like scientists, people need to generate and select the best out of many possible and plausible representations.
The flexibility in how people generate representations is demonstrated both by observations that their representations of the same environment can show considerable differences (e.g., Gaissmaier & Schooler Reference Gaissmaier and Schooler2008; Schulze & Newell Reference Schulze and Newell2016), and by observations that such representations can be improved on (e.g., Szollosi et al. Reference Szollosi, Liang, Konstantinidis, Donkin and Newell2019). The potential explanations that (a) these are in fact not new or improved representations, but result from the misapplication of strategies that evolved in a different evolutionary milieu; or (b) that people assess all potentially relevant features of the environment are both unsatisfactory. The former increases the flexibility of the model to an extent that there is almost nothing it cannot account for. The latter is computationally impossible.
This is not to say that people do not rely on heuristics to generate representations of the environment. Our argument is only that we should appreciate the flexibility of this generation process in our models, instead of substituting it with fixed representations based on flexible assumptions of the researcher. A promising way for such investigations would be the use of simple yet diagnostic manipulations of purportedly relevant features of the environment followed by thorough probing of people's knowledge about these features (e.g., Tran et al. Reference Tran, Vul and Pashler2017; also see, Newell & Shanks Reference Newell and Shanks2014).
The usefulness of resource-rational analysis hinges on the assumed similarity in representation generation between cash registers and people, because only under such conditions can the purpose of the system be clearly defined. We argued against this assumption: people seem capable of generating representations of anything, whereas cash registers can only represent things that they were programmed to represent. This difference leads to a radically different computational level question about cognition. Instead of asking what the purpose of people's cognitive mechanisms are in terms of prototypical evolutionary or learning environments, we could ask what purpose they serve in achieving the universality of representation generation. Such universality makes characterising the function of human cognition elusive, because with every new representation of a problem, the person generates a new function. Theories of human cognition need to clarify the development of purpose not merely presume its existence.