The target article by Bentley et al. provides an innovative conceptual map of human decision-making in a social context. It may well provide a valuable guide to researchers who are beginning to analyse “big data”, such as aggregate online purchasing decisions or peer-to-peer social interactions. I see their scheme as complementary to calls from myself and others (Gintis Reference Gintis2007; Mesoudi Reference Mesoudi2011; Mesoudi et al. Reference Mesoudi, Whiten and Laland2006; Richerson & Boyd Reference Richerson and Boyd2005) to restructure the social and behavioral sciences around an evolutionary framework. Evolutionary “population thinking” concerns the exact problem that is addressed throughout the target article: how individual-level processes aggregate to form population-level patterns. In biology, the individual-level processes are natural selection, genetic mutation, Mendelian inheritance, and so on, and the population-level patterns include adaptation, speciation, adaptive radiation, serial founder effects, etc. Several decades of research has identified equivalent (but often different) individual-level processes in cultural evolution (Boyd & Richerson Reference Boyd and Richerson1985; Cavalli-Sforza & Feldman Reference Cavalli-Sforza and Feldman1981; Mesoudi Reference Mesoudi2011), such as conformist- or prestige-based social learning biases (Boyd & Richerson Reference Boyd and Richerson1985), or the non-random generation of new cultural variation according to content-based inductive biases (Griffiths et al. Reference Griffiths, Kalish and Lewandowsky2008). Major advances in the social sciences can be made by borrowing tools from evolutionary biology to both explore the population-level consequences of these individual-level biases (e.g., population-genetic-style mathematical models; see Bentley et al. Reference Bentley, Hahn and Shennan2004), and quantitatively identify and measure those population-level patterns in real cultural datasets (e.g., phylogenetic methods; see O'Brien et al. Reference O'Brien, Darwent and Lyman2001).
The conceptual map presented in Figure 1 of the target article similarly links individual-level decisions (the degree to which individuals rely on social or asocial learning, represented by the east–west axis) made within different environments (the transparency in payoffs, represented by the north–south dimension) to different population-level patterns, such as the popularity distributions shown in Figure 2. While recognising the heuristic value in such a simple conceptual map, I caution that these particular dimensions may over-simplify and obscure some key issues, which the cultural evolution literature has identified in recent years as being particularly important for understanding cultural change.
First, regarding the east–west axis, it seems problematic to treat all social learning as equivalent, or at least as having broadly similar population-level consequences. Different social learning biases, such as the aforementioned conformist, prestige and inductive biases, may have very different population-level signatures. Models suggest that prestige bias can generate a runaway process towards maladaptively extreme values, while conformity generates particularly pronounced within-group behavioral homogeneity (Boyd & Richerson Reference Boyd and Richerson1985). Culturally driven copycat-suicide clusters require a particular combination of social learning processes in order to occur, primarily the rapid one-to-many transmission characteristic of the mass media plus a celebrity-driven prestige bias (Mesoudi Reference Mesoudi2009). Even restricting ourselves to the popularity distributions that Bentley et al. focus on, Mesoudi and Lycett (Reference Mesoudi and Lycett2009) showed that conformity and anti-conformity have very different consequences on population-level frequency distributions of discrete traits such as first names, with conformity creating a “winner-take-all” distribution where popular traits are made even more popular, and anti-conformity favoring traits of intermediate frequency. In sum, knowing that social (as opposed to individual) learning is at work might be useful, but knowing what kind of social learning is operating seems to be crucial, too.
Second, the north–south “transparent-opaque” dimension appears to conflate feedback error in agents’ decisions with the actual payoff structure that underlies different decisions. The north–south dimension is said to represent “the extent to which there is a transparent correspondence between an individual's decision and the consequences (costs and payoffs) of that decision” (target article, sect. 2, para. 1). In other words, it concerns feedback error: high feedback error equals opaque decision-making, while low feedback error equals transparent decision-making. This is captured formally in the b
t
parameter of equation 1. Yet this does not address the actual payoff functions underlying different choices (denoted by the function U in the equation, but unaddressed in the map).
With respect to actual payoffs, there are several possibilities: There may be a single objectively best option and many bad options, or there may be several equally good options, or there may be no functional correspondence between choice and payoff whatsoever. In adaptive landscape terms (Wright Reference Wright1932), these correspond to a unimodal, a multimodal (or rugged), and a flat landscape, respectively. The shape of this underlying adaptive landscape is logically independent to how well that payoff structure can be perceived by agents (i.e., the vertical transparency-opaqueness dimension).
I would argue that one cannot understand the consequences of transparent versus opaque feedback error without also considering the actual shape of the underlying adaptive landscape. Opaque feedback in a flat (neutral) landscape will be unproblematic, because all options are equivalent and feedback error is unimportant. However, opaque feedback in a rugged landscape will be very problematic, given the need to find one of a small number of fitness peaks. Conversely, perfectly transparent feedback may be problematic in a rugged landscape because it may lead learners to locally optimal but globally sub-optimal peaks/decisions, whereas the error intrinsic in slightly opaque feedback might lead learners, by chance, off their sub-optimal peak and onto a higher peak elsewhere in the landscape.
Experiments and models show that the shape of the adaptive landscape can significantly affect both people's choices and the aggregate outcome of those choices, quite independently of feedback error (Mesoudi Reference Mesoudi2008; Mesoudi & O'Brien Reference Mesoudi and O'Brien2008a; Reference Mesoudi and O'Brien2008b). Yet Bentley et al. appear to conflate these two distinct dimensions. For example, the neutral models discussed in section 2.3.1 (and analysed in Bentley et al. Reference Bentley, Hahn and Shennan2004) surely concern the case where the actual payoffs of all possible choices are equivalent, rather than where payoffs are opaque.
Naturally, all heuristic schemes such as the one presented by Bentley et al. are simplifications, and their value lies in that simplicity, as researchers grapple with the enormous datasets generated in the modern age. At the same time, oversimplification can sometimes lead to the wrong answer. I suspect that distinguishing between different social learning biases, and considering payoff structure as well as feedback error, might be crucial in avoiding those wrong answers.
You have
Access
The target article by Bentley et al. provides an innovative conceptual map of human decision-making in a social context. It may well provide a valuable guide to researchers who are beginning to analyse “big data”, such as aggregate online purchasing decisions or peer-to-peer social interactions. I see their scheme as complementary to calls from myself and others (Gintis Reference Gintis2007; Mesoudi Reference Mesoudi2011; Mesoudi et al. Reference Mesoudi, Whiten and Laland2006; Richerson & Boyd Reference Richerson and Boyd2005) to restructure the social and behavioral sciences around an evolutionary framework. Evolutionary “population thinking” concerns the exact problem that is addressed throughout the target article: how individual-level processes aggregate to form population-level patterns. In biology, the individual-level processes are natural selection, genetic mutation, Mendelian inheritance, and so on, and the population-level patterns include adaptation, speciation, adaptive radiation, serial founder effects, etc. Several decades of research has identified equivalent (but often different) individual-level processes in cultural evolution (Boyd & Richerson Reference Boyd and Richerson1985; Cavalli-Sforza & Feldman Reference Cavalli-Sforza and Feldman1981; Mesoudi Reference Mesoudi2011), such as conformist- or prestige-based social learning biases (Boyd & Richerson Reference Boyd and Richerson1985), or the non-random generation of new cultural variation according to content-based inductive biases (Griffiths et al. Reference Griffiths, Kalish and Lewandowsky2008). Major advances in the social sciences can be made by borrowing tools from evolutionary biology to both explore the population-level consequences of these individual-level biases (e.g., population-genetic-style mathematical models; see Bentley et al. Reference Bentley, Hahn and Shennan2004), and quantitatively identify and measure those population-level patterns in real cultural datasets (e.g., phylogenetic methods; see O'Brien et al. Reference O'Brien, Darwent and Lyman2001).
The conceptual map presented in Figure 1 of the target article similarly links individual-level decisions (the degree to which individuals rely on social or asocial learning, represented by the east–west axis) made within different environments (the transparency in payoffs, represented by the north–south dimension) to different population-level patterns, such as the popularity distributions shown in Figure 2. While recognising the heuristic value in such a simple conceptual map, I caution that these particular dimensions may over-simplify and obscure some key issues, which the cultural evolution literature has identified in recent years as being particularly important for understanding cultural change.
First, regarding the east–west axis, it seems problematic to treat all social learning as equivalent, or at least as having broadly similar population-level consequences. Different social learning biases, such as the aforementioned conformist, prestige and inductive biases, may have very different population-level signatures. Models suggest that prestige bias can generate a runaway process towards maladaptively extreme values, while conformity generates particularly pronounced within-group behavioral homogeneity (Boyd & Richerson Reference Boyd and Richerson1985). Culturally driven copycat-suicide clusters require a particular combination of social learning processes in order to occur, primarily the rapid one-to-many transmission characteristic of the mass media plus a celebrity-driven prestige bias (Mesoudi Reference Mesoudi2009). Even restricting ourselves to the popularity distributions that Bentley et al. focus on, Mesoudi and Lycett (Reference Mesoudi and Lycett2009) showed that conformity and anti-conformity have very different consequences on population-level frequency distributions of discrete traits such as first names, with conformity creating a “winner-take-all” distribution where popular traits are made even more popular, and anti-conformity favoring traits of intermediate frequency. In sum, knowing that social (as opposed to individual) learning is at work might be useful, but knowing what kind of social learning is operating seems to be crucial, too.
Second, the north–south “transparent-opaque” dimension appears to conflate feedback error in agents’ decisions with the actual payoff structure that underlies different decisions. The north–south dimension is said to represent “the extent to which there is a transparent correspondence between an individual's decision and the consequences (costs and payoffs) of that decision” (target article, sect. 2, para. 1). In other words, it concerns feedback error: high feedback error equals opaque decision-making, while low feedback error equals transparent decision-making. This is captured formally in the b t parameter of equation 1. Yet this does not address the actual payoff functions underlying different choices (denoted by the function U in the equation, but unaddressed in the map).
With respect to actual payoffs, there are several possibilities: There may be a single objectively best option and many bad options, or there may be several equally good options, or there may be no functional correspondence between choice and payoff whatsoever. In adaptive landscape terms (Wright Reference Wright1932), these correspond to a unimodal, a multimodal (or rugged), and a flat landscape, respectively. The shape of this underlying adaptive landscape is logically independent to how well that payoff structure can be perceived by agents (i.e., the vertical transparency-opaqueness dimension).
I would argue that one cannot understand the consequences of transparent versus opaque feedback error without also considering the actual shape of the underlying adaptive landscape. Opaque feedback in a flat (neutral) landscape will be unproblematic, because all options are equivalent and feedback error is unimportant. However, opaque feedback in a rugged landscape will be very problematic, given the need to find one of a small number of fitness peaks. Conversely, perfectly transparent feedback may be problematic in a rugged landscape because it may lead learners to locally optimal but globally sub-optimal peaks/decisions, whereas the error intrinsic in slightly opaque feedback might lead learners, by chance, off their sub-optimal peak and onto a higher peak elsewhere in the landscape.
Experiments and models show that the shape of the adaptive landscape can significantly affect both people's choices and the aggregate outcome of those choices, quite independently of feedback error (Mesoudi Reference Mesoudi2008; Mesoudi & O'Brien Reference Mesoudi and O'Brien2008a; Reference Mesoudi and O'Brien2008b). Yet Bentley et al. appear to conflate these two distinct dimensions. For example, the neutral models discussed in section 2.3.1 (and analysed in Bentley et al. Reference Bentley, Hahn and Shennan2004) surely concern the case where the actual payoffs of all possible choices are equivalent, rather than where payoffs are opaque.
Naturally, all heuristic schemes such as the one presented by Bentley et al. are simplifications, and their value lies in that simplicity, as researchers grapple with the enormous datasets generated in the modern age. At the same time, oversimplification can sometimes lead to the wrong answer. I suspect that distinguishing between different social learning biases, and considering payoff structure as well as feedback error, might be crucial in avoiding those wrong answers.