Robert Axelrod on the evolution of ethnocentrism

The evolution of ethnocentrism (pdf):

Ethnocentrism is a nearly universal syndrome of attitudes and behaviors, typically including in-group favoritism. Empirical evidence suggests that a predisposition to favor in-groups can be easily triggered by even arbitrary group distinctions and that preferential cooperation within groups occurs even when it is individually costly. The authors study the emergence and robustness of ethnocentric behaviors of in-group favoritism, using an agent-based evolutionary model. They show that such behaviors can become wide spread under a broad range of conditions and can support very high levels of cooperation, even in one move prisoner's dilemma games. When cooperation is especially costly to individuals, the authors show how ethnocentrism itself can be necessary to sustain cooperation.

Ethnocentrism is a nearly universal syndrome of discriminatory attitudes and behaviors (Sumner 1906; Le Vine and Campbell 1972). The attitudes include seeing one's own group (the in-group) as virtuous and superior, one's own standards of value as uni versal, and out-groups as contemptible and inferior. Behaviors associated with ethno centrism include cooperative relations within the group and the absence of cooperative relations with out-groups (LeVine and Campbell 1972). Ethnocentric behaviors are based on group boundaries that are typically defined by one or more observable characteristics (such as language, accent, physical features, or religion) regarded as indicating common descent (Sumner 1906; Hirschfeld 1996; Kurzban, Tooby, and Cosmides 2001). Such behaviors often also have a strong territorial component (Sumner 1906). Ethnocentrism has been implicated not only in ethnic conflict (Brewer 1979; Chirot and Seligman 2001), instability of democratic institutions (Rabushka and Shepsle 1972), and war (van der Dennen 1995) but also in consumer choice (Klein and Ettenson 1999) and voting (Kinder 1998). Although ethnocentrism is sometimes used to refer to a wide range of discriminatory behaviors, we will focus on ethnocentric behavior defined as in-group favoritism.

Ethnocentrism is generally thought to involve substantial cognitive ability in indi viduals (Sumner 1906; Simmel 1955; Sherif and Sherif 1956; Sherif 1966; LeVine and Campbell 1972; Hewstone, Rubin, and Willis 2002) and to be based on complex social and cultural inputs. While such factors certainly play a role in much ethnocentric behavior, extensive empirical evidence from psychology suggests the prevalence of a strong individual predisposition toward bias in favor of in-groups, which can be observed even when cognition is minimal and social input very abstract. [. . .]

The main result of the simulation is that the ethnocentric strategy becomes common even though, unlike previous models,3 favoritism toward similar others is not built into the model. In the final 100 periods of ten 2,000-period runs, 76 percent of the agents have the ethnocentric strategy, compared to 25 percent if selection had been neutral (Table 1, row a). This result shows that in-group favoritism based on simple tags and local interactions can overcome egoism and dominate a population even in the absence of reciprocity and reputation and even when "cheaters" need to be suppressed. Not only is ethnocentrism the dominant strategy, but cooperation (donation) is also the dominant behavioral choice: fully 74 percent of interactions are cooperative (Table 1, row a). Cooperation is common because the dominance of ethnocentric strategies is combined with a tendency for neighbors to have the same tag.

The emergence and dominance of the ethnocentric strategy is not a "knife-edge" phenomenon. In fact, its dominance is robust under a wide range of parameters and variations in the model. When any of the following parameters are either halved or doubled, at least two-thirds of strategies are ethnocentric: cost of helping, lattice width, number of groups, immigration rate, mutation rate, and duration of the run (see the sensitivity analysis in Table 1). The ethnocentric strategy becomes just as dominant even when the simulation starts with a full lattice consisting only of egoists, and no immigration is allowed. Another check for robustness is a variant of the model in which an agent can distinguish all four colors, rather than just distinguish ing between its own color and all other colors. Again, the results are very similar, with 80 percent ethnocentric strategies. Surprisingly, the results are also not very sensitive to the possibility that an agent will occasionally misperceive whether the ther agent in the interaction has the same color. Even when agents make this mis take 10 percent of the time, the population evolves to be more than two-thirds ethnocentric. This resistance of in-group favoritism to noise is quite a contrast to studies of reciprocity in the iterated prisoner's dilemma. The tit-for-tat strategy, for example, requires the addition of generosity or contrition to be effective in the face of even rare misperceptions (Molander 1985; Wu and Axelrod 1995).

Examining the dynamics of the model reveals how the ethnocentric strategy becomes so common and how "cheaters" are suppressed by ethnocentrics of a dif ferent color. In the early periods of a run, the scattered immigrants create regions of similar agents (Figure la). Colonies of those willing to cooperate with their own color will tend to grow faster, but over time, they face free riding by egoists who arise by mutation. Egoists who free ride cannot be suppressed by ethnocentrics of the same color and therefore tend to erode cooperative regions. Once the space is nearly full, another dynamic is added as regions with different attributes expand until they are adjacent to each other. These dynamics can be analyzed in terms of regions of contiguous agents having the same color and strategy (Figure lb). The most important aspect of regional dynamics is that an ethnocentric region will tend to expand at the expense of a region of a different color using any one of the other three strategies (Figure 2). In this way, free riding is controlled?egoists of any one color are suppressed by ethnocentric agents of different colors.

A remarkable result is that the ability to discriminate between the in-group and the out-groups can actually promote cooperation. As long as agents can distinguish their own color from other colors, even doubling the cost of cooperation sustains a cooperation level of 56 percent. However, when agents are unable to distinguish their own color from others, cooperation in the doubled-cost case falls to 14 percent. Therefore, as the cost of giving help increases, the ability to distinguish between in group and out-group members can be essential for the maintenance of cooperation in "austere" environments. In fact, the ability to distinguish between groups can be regarded as a basis for social capital within a group (Coleman 1990; Putnam 2000).

[RA Hammond, R Axelrod. The evolution of ethnocentrism. Journal of Conflict Resolution, 2006.]

Altruism via kin-selection strategies that rely on arbitrary tags with which they coevolve (pdf):

We show with an evolutionary model how contingent altruism can be sustained even when arbitrary heritable indicators of relatedness, called ‘‘tags’’, coevolve with the strategies gov- erning behavior. Discrimination based on tags is not assumed, but rather evolves endogenously in a viscous population (i.e., local reproduction and local interaction) and is selected for even when phenotypic matching is very coarse-grained. We also show how to extend Hamilton’s rule to establish the conditions under which kin recognition can support discrim- inating altruism even when coevolution causes the reliability of indicators of relatedness to vary with each individual’s evolving social environment. [. . .]

The resulting agent-based model is based on a model previously developed to study ethnocentrism in humans (Ax- elrod and Hammond 2003). The present model is not meant to be a literal representation of biological processes. Instead, our model is designed to illuminate the consequences of the fact that kin discrimination typically entails coevolution of three things: the strategies governing behavior, the reliability of the tags on which the behavior may be conditioned, and the population structure that determines who interacts with whom. [. . .]

The algebraic method above is the first published analysis of selection for kin recognition with simultaneous variation at the indicator and altruistic loci. This method helps us un- derstand the conditions under which kin recognition can sup- port discriminating altruism even when the reliability of in- dicators of kinship depends on the individual’s social envi- ronment.

The value of being able to distinguish tags can be under- stood in terms of inclusive fitness theory that takes into ac- count the degree of relatedness between two agents (Hamilton 1964; Lacy and Sherman 1983; Riolo et al. 2001). While proximity alone can be an indication of relatedness, being able to distinguish among heritable tags, as in the armpit effect (Dawkins 1982; Hauber and Sherman 2000; Hauber et al. 2000; Mateo and Johnson 2000; Isles et al. 2001), allows a still better indication of relatedness, for example among sessile cnidarians (Grosberg and Quinn 1989; Grafen 1990). The discriminatory abilities required for the armpit effect are likely to be widespread. The self-recognition required for multicellularity provides them from intimate contact, and the need to distinguish conspecifics for mating provides them more generally for animals. In both cases, a hardwired com- parison known as the green beard effect (Hamilton 1964; Dawkins 1976; Haig 1996; Grafen 1998; Keller and Ross 1998) would seriously slow evolution and make speciation almost impossible.

Viscosity is ubiquitous because few populations complete- ly mix from one generation to the next. Hamilton (1964) believed that simple viscosity was a widespread sufficient cause of fairly weak altruism, and various models have found that viscosity can indeed foster cooperation (Getty 1987; Pol- lock 1989; Nowak and May 1992; Nakamaru et al. 1997). However, this general claim is now considered doubtful. The balance between increased relatedness and increased com- petition between neighbors may tilt toward or away from cooperation (Taylor 1992; Wilson et al. 1992; West et al. 2002). Taylor and Irwin (2000) have suggested that with overlapping generations, and with altruism dispensed as ben- efits to fecundity, there is a tendency for population viscosity to support altruism. The 15.6% cooperation found in our model with one tag is on the one hand more than zero, sup- porting Taylor and Irwin, but on the other hand is rather limited. Adding observable tags shows that proximity can sustain cooperation based on contingent altruism, even if the very correlation of tags and relatedness evolves. By putting both the matching and the altruism under explicit genetic control, the model shows how altruism conditional on heritable tags can evolve despite substantial costs of cooperation. Thus, the present model, which combines viscosity, the armpit effect, and endogenous use of discrimination in a genetically explicit way, creates a very general expectation of widespread, and not necessarily weak, conditional altruism in nature.

[R Axelrod, RA Hammond, A Grafen. Altruism via kin-selection strategies that rely on arbitrary tags with which they coevolve.]

And some related work:

Recent agent-based computer simulations suggest that ethnocentrism, often thought to rely on complex social cognition and learning, may have arisen through biological evolution. From a random start, ethnocentric strategies dominate other possible strategies (selfish, traitorous, and humanitarian) based on cooperation or non-cooperation with in-group and out-group agents. Here we show that ethnocentrism eventually overcomes its closest competitor, humanitarianism, by exploiting humanitarian cooperation across group boundaries as world population saturates. Selfish and traitorous strategies are self-limiting because such agents do not cooperate with agents sharing the same genes. Traitorous strategies fare even worse than selfish ones because traitors are exploited by ethnocentrics across group boundaries in the same manner as humanitarians are, via unreciprocated cooperation. By tracking evolution across time, we find individual differences between evolving worlds in terms of early humanitarian competition with ethnocentrism, including early stages of humanitarian dominance. Our evidence indicates that such variation, in terms of differences between humanitarian and ethnocentric agents, is normally distributed and due to early, rather than later, stochastic differences in immigrant strategies.

[Max Hartshorn, Artem Kaznatcheev and Thomas Shultz. The Evolutionary Dominance of Ethnocentric Cooperation. Journal of Artificial Societies and Social Simulation 16 (3) 7]

Related: Hamilton on inclusive fitness and social behavior in humans

5 comments:

Anonymous said...

Great post!

RCB said...

Pretty interesting. First paper would presumably suggest that selection would favor ethnocentrism based on quickly evolving cultural, but not genetic, characteristics. (But maybe highly mutable genes - a mutation rate of .005 is pretty extreme.)

One minor concern: I think the cultural transmission model could be explored here. Culture is modeled to behave like highly mutable genes, with parental transmission. But most gene-culture coevolution models allow for other kinds of transmission pathways. If you introduce horizontal cultural transmission (common in gene-culture coevolution models), for example, then these cultural tags will spread among non-kin, and then the system would break down, I imagine. But not entirely, perhaps.

Anyway, cool simulation.

RCB said...

The first model (Hammond and Axelrod) has been stuck in the back of my head for a while. So I rebuilt it using R and C++. I was able to replicate the main results: that there is a wide range under which "tag-based" altruism evolves, under the model's particular assumptions.

But adding two of realistic alterations (in my opinion) causes this result to break down.

(1) The model assumes asexual reproduction. Adding sexual reproduction with recombination (as for humans) causes the tags to become much less associated with other loci. That means that tags become a much less reliable indicator of recent decent (and therefore loci at other genes), so tag-based altruism becomes untenable.
(2) The model apparently has no dispersal. That's obviously inaccurate. In fact, under sexual reproduction, no dispersal in this model implies extremely strong inbreeding (e.g., 2/8 chance of mating with your parent, since 2 of your 8 neighbors are parents!). Adding dispersal greatly reduces the scope for tag-based altruism.

So, I'm much less swayed by this paper now.

n/a said...

RCB,

I'm not sure what sort of dispersal you have in mind, but the model clearly does include dispersal in the simple meaning of the term. If you mean some sort of random, long-distance dispersal of lone individuals, a model of that sort would have little applicability to humans.

Here's a group that added sexual reproduction to a similar model:

Evolution of ethnocentrism on undirected and directed Barab├ísi–Albert networks

F.W.S. Lima 1 , Tarik Hadzibeganovic 2,3 and Dietrich Stauffer 4

Abstract. Using Monte Carlo simulations, we study the evolution of contigent co-
operation and ethnocentrism in the one-move game. Interactions and reproduction
among computational agents are simulated on undirected and directed Barabási-
Albert (BA) networks. We first replicate the Hammond-Axelrod model of in-
group favoritism on a square lattice and then generalize this model on undirected
and directed BA networks for both asexual and sexual reproduction cases. Our
simulations demonstrate that irrespective of the mode of reproduction, ethnocentric strategy becomes common even though cooperation is individually costly and
mechanisms such as reciprocity or conformity are absent
. Moreover, our results
indicate that the spread of favoritism toward similar others highly depends on the
network topology and the associated heterogeneity of the studied population.

http://arxiv.org/pdf/0905.2672

RCB said...

Thanks.

Re: dispersal. In the Hammond and Axelrod model (http://www.artisresearch.com/articles/Axelrod_Evolution_of_Ethnocentrism.pdf), there is no dispersal: offspring are born in a square adjacent to the parent, and they stay there indefinitely. This is not dispersal, unless you define dispersal as "not inhabiting the exact same physical space as your parent." If you add sexual reproduction, and assume that reproduction occurs among adjacent squares (as I did), then the result is that most matings occur with parents, siblings, and cousins. That's just not how it is. It also implies that virtually all social interactions occur among these very close kin, which again appears to be not true. (see links below.)

I added dispersal by allowing that people "leave home" with some probability. If they do leave home, they randomly switch places with someone else on the board. Of course, in reality, short-distance dispersal should be more common than long, so there is room for improvement there. But this is only a 50x50 grid - even "long range" dispersal isn't that long (50 people away - like moving one or two villages away). And even a little bit of dispersal kills the results, it seems.

Re: sexual reproduction: I checked with my model, and it does appears that sexual reproduction alone (without dispersal added) doesn't have a big effect. So it may be that dispersal is what really kills the results - or the combination of the two. (I am slightly bothered, btw, that the words "recombination" and "inbreeding" never show up in the Lima paper.)

Overall, I'm not convinced that the 2D grid of individuals is a great way to model human groups. Probably a better structure would be a grid of groups (bands, tribes, whatever), rather than individuals, in which people can interact with every tag-carrying member of their group. Then have some dispersal between groups. With that model, you wouldn't get extremely small social/mating networks (like the 8 implied by Hammond + Axelrod), which are untenable for most human populations (papers below).

Some useful recent research on the fluidity of extant hunter-gatherer populations.
http://anthro.vancouver.wsu.edu/media/PDF/Science_final.pdf
http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0102806