In their target article, Baumeister et al. make a case for the claim that group performance depends in crucial ways on the degree to which group members identify with the group as well as on the degree of individual differentiation among group members. Baumeister et al. muster an impressive amount of evidence for this claim, where the evidence comes mostly from social psychology and personality research.
This commentary draws attention to the fact that there is still a large and important body of literature that is highly relevant to the claim made in the target article but that is entirely neglected by Baumeister et al. Computer scientists, economists, cognitive scientists, and others have in recent years systematically studied group performance by means of computational modeling, in particular agent-based simulations. Although it is generally acknowledged that the models used in these studies are idealized in a number of ways, it is noteworthy that some of the features that Baumeister et al. identify as contributing to group performance have also been identified as such in simulation-based studies of group behavior. The interesting point is that, whereas Baumeister et al. attribute those features to the attitudes of group members toward the group, or the moral behavior of those members, or – most important for their main claim – the extent to which members are identifiable within the group, in the computational modeling literature, these features arise partly from very different assumptions: for example, concerning the (mathematically characterizable) structure of the group, or the spread of opinions within the group, or sometimes even just the size of the group.
There is a vast literature on agent-based simulations, in which a variety of models have been investigated. Here, I focus on one particular type of models – the so-called Hegselmann-Krause (HK) models and their variants – and some of the results the study of these models has led to.
In their original form, the HK models simulate groups of agents who try to determine the unknown value of some parameter by exchanging information with other agents in the group. In the simplest model, the agents repeatedly update their opinions about the value of the parameter by averaging the opinions of the agents that are in their Bounded Confidence Interval (BCI), where agent A is in agent B's BCI precisely if A's opinion is “close enough” to B's. (In their own studies, Hegselmann & Krause investigated systematically different bounds on the confidence interval.) In a more interesting model, agents also receive evidence directly from the world about the value of the parameter they aim to determine, and they update their opinions by “mixing,” in a specific way, that evidence and the opinions of the agents in their BCI. (For details, see Hegselmann & Krause Reference Hegselmann and Krause2002; Reference Hegselmann and Krause2005; Reference Hegselmann and Krause2006; Reference Hegselmann and Krause2009.)
In recent years, a number of extensions of the HK-models, in particular of the second one, have been proposed. For example, extensions have been studied in which communities of agents hold sets of logically related beliefs, rather than a single opinion on the value of a parameter (Riegler & Douven Reference Riegler and Douven2009; Wenmackers et al. Reference Wenmackers, Vanpoucke and Douven2012; Reference Wenmackers, Vanpoucke and Douven2014). And Douven and Wenmackers (Reference Douven and Wenmackersin press) present results concerning a version of the second HK-model in which communities of agents update probabilities, where the focus of Douven and Wenmackers' study is the comparative efficacy of different update rules.
Some of the results obtained in the above and related studies bear directly on the topic of the target article. To mention a few: (1) Sometimes the reluctance of agents to share information with anyone but those whose opinions are extremely close to their own can be beneficial for the group as a whole (Hegselmann & Krause Reference Hegselmann and Krause2002; Reference Hegselmann and Krause2006). (2) Free riding, in the sense of agents' ignoring evidence coming directly from the world, is tolerable to a surprisingly high degree: As long as some agents do take that evidence into account, all agents will arrive at holding a true belief (or true beliefs, in some models), and the group as a whole will not be significantly slowed down by the free riders (Hegselmann & Krause Reference Hegselmann and Krause2002; Reference Hegselmann and Krause2006). (3) Giving greater weight to the opinions of experts does not always pay off for the group as a whole (Douven & Riegler Reference Douven and Riegler2010). (4) If the evidence the agents receive is noisy (as much real-world evidence is), then the average agent's opinion may converge faster on the truth if the agents do not communicate their opinions with other agents; but it may, eventually, approach the truth more closely if the agents do communicate their opinions (Douven Reference Douven2010).
For any of these phenomena to occur, it is immaterial whether the agents can be held accountable for either their opinions or their communicative behavior, or indeed whether they are identifiable for other agents at all. This is not to suggest that the kind of explanations for group performance that Baumeister et al. canvass – at least insofar as they pertain to information use (which they distinguish from productive achievement) – are wrong or superfluous. What it does suggest is that, for at least some of the phenomena Baumeister et al. seek to explain, there is a problem of overdetermination: These phenomena may be a result of the attitudes that group members hold toward each other (etc.), but they may also arise out of much simpler facts, completely unrelated to whether agents care about how well their group does or about how they are perceived by other members of the group.
It is reasonable to hold that, in reality, both types of factors will play a role. This should give computational modelers reason to try to incorporate in their simulations the kind of factors that Baumeister et al. discuss. At the same time, the findings from the literature on agent-based simulations should give social psychologists reason to investigate which part of group performance is best explained by broadly moral considerations of the group members and which part by structural constraints on information exchange that may not be morally evaluable.
You have
Access
In their target article, Baumeister et al. make a case for the claim that group performance depends in crucial ways on the degree to which group members identify with the group as well as on the degree of individual differentiation among group members. Baumeister et al. muster an impressive amount of evidence for this claim, where the evidence comes mostly from social psychology and personality research.
This commentary draws attention to the fact that there is still a large and important body of literature that is highly relevant to the claim made in the target article but that is entirely neglected by Baumeister et al. Computer scientists, economists, cognitive scientists, and others have in recent years systematically studied group performance by means of computational modeling, in particular agent-based simulations. Although it is generally acknowledged that the models used in these studies are idealized in a number of ways, it is noteworthy that some of the features that Baumeister et al. identify as contributing to group performance have also been identified as such in simulation-based studies of group behavior. The interesting point is that, whereas Baumeister et al. attribute those features to the attitudes of group members toward the group, or the moral behavior of those members, or – most important for their main claim – the extent to which members are identifiable within the group, in the computational modeling literature, these features arise partly from very different assumptions: for example, concerning the (mathematically characterizable) structure of the group, or the spread of opinions within the group, or sometimes even just the size of the group.
There is a vast literature on agent-based simulations, in which a variety of models have been investigated. Here, I focus on one particular type of models – the so-called Hegselmann-Krause (HK) models and their variants – and some of the results the study of these models has led to.
In their original form, the HK models simulate groups of agents who try to determine the unknown value of some parameter by exchanging information with other agents in the group. In the simplest model, the agents repeatedly update their opinions about the value of the parameter by averaging the opinions of the agents that are in their Bounded Confidence Interval (BCI), where agent A is in agent B's BCI precisely if A's opinion is “close enough” to B's. (In their own studies, Hegselmann & Krause investigated systematically different bounds on the confidence interval.) In a more interesting model, agents also receive evidence directly from the world about the value of the parameter they aim to determine, and they update their opinions by “mixing,” in a specific way, that evidence and the opinions of the agents in their BCI. (For details, see Hegselmann & Krause Reference Hegselmann and Krause2002; Reference Hegselmann and Krause2005; Reference Hegselmann and Krause2006; Reference Hegselmann and Krause2009.)
In recent years, a number of extensions of the HK-models, in particular of the second one, have been proposed. For example, extensions have been studied in which communities of agents hold sets of logically related beliefs, rather than a single opinion on the value of a parameter (Riegler & Douven Reference Riegler and Douven2009; Wenmackers et al. Reference Wenmackers, Vanpoucke and Douven2012; Reference Wenmackers, Vanpoucke and Douven2014). And Douven and Wenmackers (Reference Douven and Wenmackersin press) present results concerning a version of the second HK-model in which communities of agents update probabilities, where the focus of Douven and Wenmackers' study is the comparative efficacy of different update rules.
Some of the results obtained in the above and related studies bear directly on the topic of the target article. To mention a few: (1) Sometimes the reluctance of agents to share information with anyone but those whose opinions are extremely close to their own can be beneficial for the group as a whole (Hegselmann & Krause Reference Hegselmann and Krause2002; Reference Hegselmann and Krause2006). (2) Free riding, in the sense of agents' ignoring evidence coming directly from the world, is tolerable to a surprisingly high degree: As long as some agents do take that evidence into account, all agents will arrive at holding a true belief (or true beliefs, in some models), and the group as a whole will not be significantly slowed down by the free riders (Hegselmann & Krause Reference Hegselmann and Krause2002; Reference Hegselmann and Krause2006). (3) Giving greater weight to the opinions of experts does not always pay off for the group as a whole (Douven & Riegler Reference Douven and Riegler2010). (4) If the evidence the agents receive is noisy (as much real-world evidence is), then the average agent's opinion may converge faster on the truth if the agents do not communicate their opinions with other agents; but it may, eventually, approach the truth more closely if the agents do communicate their opinions (Douven Reference Douven2010).
For any of these phenomena to occur, it is immaterial whether the agents can be held accountable for either their opinions or their communicative behavior, or indeed whether they are identifiable for other agents at all. This is not to suggest that the kind of explanations for group performance that Baumeister et al. canvass – at least insofar as they pertain to information use (which they distinguish from productive achievement) – are wrong or superfluous. What it does suggest is that, for at least some of the phenomena Baumeister et al. seek to explain, there is a problem of overdetermination: These phenomena may be a result of the attitudes that group members hold toward each other (etc.), but they may also arise out of much simpler facts, completely unrelated to whether agents care about how well their group does or about how they are perceived by other members of the group.
It is reasonable to hold that, in reality, both types of factors will play a role. This should give computational modelers reason to try to incorporate in their simulations the kind of factors that Baumeister et al. discuss. At the same time, the findings from the literature on agent-based simulations should give social psychologists reason to investigate which part of group performance is best explained by broadly moral considerations of the group members and which part by structural constraints on information exchange that may not be morally evaluable.