The notion of transparency is among the few most fundamental concepts in Bentley et al.’s target article. Transparency, and its absence or opposite, define one of the two axes of the map of collective behaviour that is described and defended there. Unfortunately, it is far from clear what notion of transparency the authors have in mind.
The fact that Bentley et al. offer an equation in which transparency occurs (as b
t
in Equation 1, sect. 2) does not clarify what they mean by transparency. Rather, it highlights the challenge they need to meet. An equation represents purely formal relationships. To make an equation empirically significant the variables need interpretations. These should include some indication of what would count as determining a value for each variable. This is as true in physics as it is in the cognitive and behavioural sciences.
One reason it isn't clear how to understand statements about transparency in the target article, is that the term itself appears in some statements that purport to explain what it means. For example, in section 2 we are told that the north–south dimension of the map graphs the “transparency or opaqueness of the decision in terms of payoff” (sect. 2, para. 1). Shortly after, in the same paragraph, we read a “more formal” statement, according to which transparency concerns the “extent to which there is a transparent correspondence between an individual's decision and the consequences (costs and payoffs) of that decision.” But, first, the more formal statement seems to say no more than the preceding one. Second, both contain the very term to be clarified.
We are also told that transparency is somehow related to payoffs. The article provides some encouragement for the view that transparency is a matter of knowing the payoff to be expected from a choice before it is made. We read, for example, that the north of the map “contains agents … who know the impact their decisions will have on them” (sect. 2.1, para. 1; emphasis added), and that going south takes us to agents who are “less able to discern differences in potential payoffs” (sect. 2, para. 1; emphasis added). Unfortunately, we're also told that in the extreme north the impact of a decision is “immediately detectable” (sect. 2, para. 4; emphasis added), which suggests transparency has something to do with experienced consequences rather than ones known in advance. These possibilities are importantly distinct. The cognitive demands of reinforcement learning and those of predicting without experience are quite different. (They correspond to Dennett's [Reference Dennett1996] distinction between “Skinnerian” and “Popperian” agents.)
The ambiguity between payoffs anticipated and experienced is mirrored in ambiguity over whether transparent payoff relations are simply known, or are the object of some kind of higher-order cognition. Sometimes the north is characterised in terms of what an agent can “discern” or “detect,” suggesting that payoffs are simply known. (Although, to the south, apparently, agents might not know the “qualifications” of those they learn from, which suggests that knowledge of matters other than payoffs might also be part of transparency.) But we're also told that agents on the northern edge “clearly understand the rationale” for new behaviours, and engage in “conscious weighing of options” (sect. 2.2, para. 1). Again, these are importantly distinct: Agents could successfully learn which behaviours are rewarded (and by how much) without being capable of means-end rationality or in any other way having a rationale for their behaviour.
What is said about the opposite of transparency raises further confusing associations. So, at the extreme south (where b
t
= 0) we encounter “total indifference,” suggesting that under conditions of full opacity agents have no preferences between options. In the deep south, “the probability of any particular choice among N
t
possible choices approaches zero (because 1/N
t
goes to zero as N
t
goes to infinity)” (sect. 2, para. 4; emphasis added). The claim here is that loss of transparency is an effect of the sheer number of options. But this seems confused. An agent could surely be entirely ignorant of the payoff difference between as few as two options (so minimal transparency occurs when N
t
is very small), and, conversely, know perfectly well that at least one option from a possibly infinite set pays handsomely.
On two occasions (in section 2) the notion of transparency is linked to a “landscape” of costs and payoffs. We are told that with northward movement, agents’ decisions will be more “attuned” with this landscape, and that the landscape can be “represent[ed] by the function U(...) of costs and payoffs” (sect. 2, para. 4). But these references to landscapes dispel little of the mist surrounding the concept of transparency.
If U(...) refers to Eq. 1, then it isn't exclusively a function about payoffs, because that includes the (presumably comprehensive) “list of covariates” (sect. 2, para. 2) that “influence choice” at a time, as well as measures of popularity (of the choice) and influence (upon the individual agent of those already opting for the choice).
Perhaps the idea is that the net individual payoff portion of U(...) could be extracted separately for each available choice. That would enable us to define the height of the “landscape” for each option (k). What we would have then would be a lookup table of payoffs and options, rather than a landscape. To organise the payoff-option pairings into a landscape, we would need to have criteria for relative proximity. (On a fitness landscape, for example, these could be provided by degrees of genetic similarity.) But what might provide homogenous criteria for degrees of proximity in the space of choices, so that the payoffs can be organised into a landscape? What topographic properties might the resultant landscape have? How are other components of U(...), such as relative popularity, to be represented on the landscape? The problem isn't that no answers to these questions are possible, it's that the target article gives no hint as to which we should have in mind.
The net result of all this is that the map defended in the target article belongs towards the (opaque) south of the very space it represents.
You have
Access
The notion of transparency is among the few most fundamental concepts in Bentley et al.’s target article. Transparency, and its absence or opposite, define one of the two axes of the map of collective behaviour that is described and defended there. Unfortunately, it is far from clear what notion of transparency the authors have in mind.
The fact that Bentley et al. offer an equation in which transparency occurs (as b t in Equation 1, sect. 2) does not clarify what they mean by transparency. Rather, it highlights the challenge they need to meet. An equation represents purely formal relationships. To make an equation empirically significant the variables need interpretations. These should include some indication of what would count as determining a value for each variable. This is as true in physics as it is in the cognitive and behavioural sciences.
One reason it isn't clear how to understand statements about transparency in the target article, is that the term itself appears in some statements that purport to explain what it means. For example, in section 2 we are told that the north–south dimension of the map graphs the “transparency or opaqueness of the decision in terms of payoff” (sect. 2, para. 1). Shortly after, in the same paragraph, we read a “more formal” statement, according to which transparency concerns the “extent to which there is a transparent correspondence between an individual's decision and the consequences (costs and payoffs) of that decision.” But, first, the more formal statement seems to say no more than the preceding one. Second, both contain the very term to be clarified.
We are also told that transparency is somehow related to payoffs. The article provides some encouragement for the view that transparency is a matter of knowing the payoff to be expected from a choice before it is made. We read, for example, that the north of the map “contains agents … who know the impact their decisions will have on them” (sect. 2.1, para. 1; emphasis added), and that going south takes us to agents who are “less able to discern differences in potential payoffs” (sect. 2, para. 1; emphasis added). Unfortunately, we're also told that in the extreme north the impact of a decision is “immediately detectable” (sect. 2, para. 4; emphasis added), which suggests transparency has something to do with experienced consequences rather than ones known in advance. These possibilities are importantly distinct. The cognitive demands of reinforcement learning and those of predicting without experience are quite different. (They correspond to Dennett's [Reference Dennett1996] distinction between “Skinnerian” and “Popperian” agents.)
The ambiguity between payoffs anticipated and experienced is mirrored in ambiguity over whether transparent payoff relations are simply known, or are the object of some kind of higher-order cognition. Sometimes the north is characterised in terms of what an agent can “discern” or “detect,” suggesting that payoffs are simply known. (Although, to the south, apparently, agents might not know the “qualifications” of those they learn from, which suggests that knowledge of matters other than payoffs might also be part of transparency.) But we're also told that agents on the northern edge “clearly understand the rationale” for new behaviours, and engage in “conscious weighing of options” (sect. 2.2, para. 1). Again, these are importantly distinct: Agents could successfully learn which behaviours are rewarded (and by how much) without being capable of means-end rationality or in any other way having a rationale for their behaviour.
What is said about the opposite of transparency raises further confusing associations. So, at the extreme south (where b t = 0) we encounter “total indifference,” suggesting that under conditions of full opacity agents have no preferences between options. In the deep south, “the probability of any particular choice among N t possible choices approaches zero (because 1/N t goes to zero as N t goes to infinity)” (sect. 2, para. 4; emphasis added). The claim here is that loss of transparency is an effect of the sheer number of options. But this seems confused. An agent could surely be entirely ignorant of the payoff difference between as few as two options (so minimal transparency occurs when N t is very small), and, conversely, know perfectly well that at least one option from a possibly infinite set pays handsomely.
On two occasions (in section 2) the notion of transparency is linked to a “landscape” of costs and payoffs. We are told that with northward movement, agents’ decisions will be more “attuned” with this landscape, and that the landscape can be “represent[ed] by the function U(...) of costs and payoffs” (sect. 2, para. 4). But these references to landscapes dispel little of the mist surrounding the concept of transparency.
If U(...) refers to Eq. 1, then it isn't exclusively a function about payoffs, because that includes the (presumably comprehensive) “list of covariates” (sect. 2, para. 2) that “influence choice” at a time, as well as measures of popularity (of the choice) and influence (upon the individual agent of those already opting for the choice).
Perhaps the idea is that the net individual payoff portion of U(...) could be extracted separately for each available choice. That would enable us to define the height of the “landscape” for each option (k). What we would have then would be a lookup table of payoffs and options, rather than a landscape. To organise the payoff-option pairings into a landscape, we would need to have criteria for relative proximity. (On a fitness landscape, for example, these could be provided by degrees of genetic similarity.) But what might provide homogenous criteria for degrees of proximity in the space of choices, so that the payoffs can be organised into a landscape? What topographic properties might the resultant landscape have? How are other components of U(...), such as relative popularity, to be represented on the landscape? The problem isn't that no answers to these questions are possible, it's that the target article gives no hint as to which we should have in mind.
The net result of all this is that the map defended in the target article belongs towards the (opaque) south of the very space it represents.