This action might not be possible to undo. Are you sure you want to continue?
Evolution of cooperation without reciprocity
Rick L. Riolo*, Michael D. Cohen² & Robert Axelrod³
* Center for the Study of Complex Systems; and ² School of Information; and ³ Gerald R. Ford School of Public Policy, University of Michigan, Ann Arbor, Michigan 48109, USA
A long-standing problem in biological and social sciences is to understand the conditions required for the emergence and maintenance of cooperation in evolving populations. For many situations, kin selection1 is an adequate explanation, although kinrecognition may still be a problem. Explanations of cooperation between non-kin include continuing interactions that provide a shadow of the future (that is, the expectation of an ongoing relationship) that can sustain reciprocity2±4, possibly supported by mechanisms to bias interactions such as embedding the agents in a two-dimensional space4±6 or other context-preserving networks7. Another explanation, indirect reciprocity8, applies when benevolence to one agent increases the chance of receiving help from others. Here we use computer simulations to show that cooperation can arise when agents donate to others who are suf®ciently similar to themselves in some arbitrary characteristic. Such a characteristic, or `tag', can be a marking, display, or other observable trait. Tag-based donation can lead to the emergence of cooperation among agents who have only rudimentary ability to detect environmental signals and, unlike models of direct3,4 or indirect reciprocity9,10, no memory of past encounters is required. Tag-based mechanisms may be useful even if an agent is unable to observe or remember others' actions. To show how tag-based strategies of helping can lead to the emergence of cooperation, we use the donor/recipient setting of ref. 10. In this setting agents are paired at random, and one agent has an opportunity to make a costly donation to the other. Unlike models of tag-based cooperation using an iterated `prisoner's dilemma' setting11±14, this setting has no continuing interaction between pairs of agents. In Nowak and Sigmund's model of `image scoring'10, an agent's decision to donate (cooperate) is based on whether the potential recipient is known to be suf®ciently generous to others. In our model, an agent's decision to donate depends only on arbitrary `tags' associated with the agents. In particular, when an agent meets another agent, it decides to help if and only if both tags are suf®ciently similar. Because agents interact with only a few randomly selected others from a moderately sized population, there is little chance that a given pair will meet again. Whether or not an agent donates does not affect the likelihood of receiving help from others. Thus there is no reciprocity, either direct or indirect. In Holland's original formulation12, arbitrary, evolving tags could facilitate selective interactions, and thereby be helpful for aggregation and boundary formation. While a tag-matching mechanism can be arbitrarily complex, we use the simplest mechanism. Each agent has two traits, a tag t [ 0; 1, and a tolerance threshold T $ 0. Initially, tags and tolerance levels are assigned to agents at random, uniformly sampled from 0; 1. (In other experiments, we started with high, T 0:5, and low, T 0:005, tolerances. Except for short initial transients, the results were not substantially different from those reported here.) In each generation, each agent acts as a potential donor with P others chosen at random, with replacement. Thus for P 3, each agent has three opportunities to donate, and on average is chosen three times as a potential recipient. An agent A donates to a potential recipient B only if B's tag is suf®ciently similar to A's tag. In particular, A donates only when B's tag is within A's tolerance threshold, TA, namely when jtA 2 tB j # T A . Thus an agent with a high T will donate to agents
NATURE | VOL 414 | 22 NOVEMBER 2001 | www.nature.com
with a wide range of tags, while an agent with a very small T will donate only to those whose tags are nearly the same as its own. If A does donate to B, A pays a cost, c, and B receives a bene®t, b. We note that even equality of tags does not make two agents more likely to interact, but if they do interact then one will donate to the other. After all agents have participated in all pairings in a generation, agents are reproduced on the basis of their score relative to others. The least ®t, median ®t, and most ®t agents have respectively 0, 1 and 2 as the expected number of their offspring. This is accomplished by comparing each agent with another randomly chosen agent, and giving an offspring to the one with the higher score. (Another interpretation of this adaptive process is learning in a ®xed population. In the learning interpretation, each agent compares itself to another agent, and adopts the other's tag and tolerance if the other's score is higher than its own.) Each offspring is subject to potential mutation which may change the offspring's tag, tolerance or both. With probability 0.1, the offspring receives a new tag with a value drawn at random in 0; 1. Also with probability 0.1, the tolerance is mutated by adding mean 0, standard deviation 0.01 gaussian noise to the old tolerance. If the new T , 0, it is set to 0. One run of the model consists of 100 agents and 30,000 generations. Each experimental condition is replicated 30 times. We ®nd that using the tag-based mechanism and adaptive processes described above, a population of agents is able rapidly to establish a substantial degree of cooperation. For example, with P 3 pairings per agent per generation, and with cost c 0:1 and bene®t b 1:0, the average donation rate was 73.6%. Figure 1 shows the dynamics of the donation rate over the ®rst 500 generations of a typical run using these parameters. The average payoff for the population at any time is proportional to the donation rate because each donation results in one agent gaining b 1:0 and another agent losing c 0:1, for a net gain to the population of 0.9. The population starts with tags and tolerances uniformly distributed, so the initial average tolerance is about 0.5, and the initial average donation rate is about 67%. Within a few generations, however, the agents with low tolerances begin to take over the population as they receive bene®ts from more tolerant agents, but they bear less of the cost because they donate to fewer others. By generation 70 in the run shown, the average tolerance is down to 0.020, and the donation rate is down to 43%. By chance there are some small groups of agents with similar tags and relatively low tolerances. As these agents prosper and reproduce, their offDonation rate (%)
75 50 25 0 0 100 200 300 400 500
Tolerance 0.03 0.02 0.01 0 0 100 200 300 Generations 400 500
Figure 1 Population dynamics for the ®rst 500 generations of a typical run. a, The donation rate. b, The average tolerance. Occasionally a mutant arises with a tag similar to most of the others, but with an unusually low tolerance. This mutant scores well by receiving donations from many, but donating to few. Its offspring quickly become numerous enough to lower the average donation rate and tolerance of the whole population. Soon their tag becomes the most common, resulting in a transition to a new dominant tag cluster. This happened at generation 226 and 356. We note that after these transitions, the average donation rate returned to its previous high level.
© 2001 Macmillan Magazines Ltd
.. We note that increasing the number of pairings of potential donors and recipients per generation increases the donation rate.... In our model....... Ten periods later the dominant cluster's relatedness increases to 97% as the members who give to the modal type without receiving donations from them are eliminated...011 0.025 0..13±17........ Pairings is the number of times per generation each agent has an opportunity to donate to a randomly encountered other.. and also to receive donations.... and then rises gradually to 79% for ten pairings................ there is only a slight selection pressure to lower their tolerance if there are no agents `nearby'... whereas the average tolerance of the population increases only slightly................... c. and defectors7............... Soon about 75±80% of the agents have tags that are so similar that they are within each other's tolerance range.. 1). except for any further mutation of tags..... In our model.. common descent has a strong in¯uence on the maintenance of cooperation. That full history is the basis of our reported averages... The intolerant agent gains ®tness because the tolerant agents donate to it....05 0........6% over the entire set of 30 runs of 30. Here. The donation rate is the percentage of such encounters in which the choosing agent cooperates.......007 0.............. the number of pairings per agent per generation is held constant at P 3... the average tolerance of a dominant cluster is much higher at its end (0......... We note that the donation rate increases markedly between P 2 and P 3........ Suppose instead that the agent adopts the better agent's tag and tolerance with probability proportional to how much better the other agent is.3....027) than its beginning (0.....2 0....... c/b.009 0.. Donation rate (%) 2.....5 0.... also affects the rate of donation............... as are tolerances..4 0............024 0.018 0...... The agents in the resulting `dominant tag cluster' have an advantage as there are more of them to help each other.... Table 2 shows how the donation rate and average tolerance are affected by varying c/b when P is held constant at P 3......... that is.6 .... Thus.. For the agents in the dominant cluster.019 0..7 79............ As expected.. but only modestly... The new dominant cluster tends to have relatively low tolerance owing to inheritance from its founder.6 73.... ... the cost/bene®t ratio c/b is altered by adjusting the cost.... while rarely giving help to others.... by establishing dominant tag clusters.... each agent compares itself to another agent..... the donation rate rapidly collapses..8 78...........000 generations each...010).. We note that increasing the cost of donating relative to the bene®ts conveyed decreases the donation rate..1 24. while it bears little cost because the smaller range of tag values to which it will donate means that it will tend to donate to fewer others.. Table 1 shows the effects of varying the number of pairings per agent per generation......... As the tolerance of a new dominant cluster increases............1 4...024 Donation rate (%) 73.... but are less sensitive than before to increases in the number of pairings....005 .... but both are heritable and subject to mutation......1 0... the average donation rate in the population falls markedly (see Fig........... The average tolerance can therefore drift upwards owing to mutation occurring near the ¯oor (base value) imposed by T $ 0... With one or two pairings... thus increasing the spread of agents with similar tags in future generations..3 0..... the amount of cooperation is less than 5%.. A potential donor agent A in a pair donates to a potential recipient B only if the distance between the tags of A and B is less than or equal to A's tolerance....... This pressure is a result of the advantage a relatively intolerant agent has in a group of more tolerant donors with more or less similar tags..4.021 0.. the rate of donation is at a high level..... When agent A donates to agent B......... and all other parameters are unchanged........ As the new cluster grows to about 75± 80% of the population.. for example...6 73.. Over time........... each one has a better chance of being found by an agent that will contribute to it. it becomes vulnerable to yet another relatively intolerant mutant with a similar tag.. This cycle corresponds to the one found in many models of the iterated `prisoner's dilemma' among conditional cooperators............ and then persists stochastically over the entire 30..... The formation of a dominant tag cluster leads to high donation rates even when averaged over the whole population... a loss of sensory discrimination in a population when there is little selection pressure to retain it. With three pairings. but the mutant's own tolerance range is small enough to prevent its donation to members of the dominant cluster. Excluding the transient period of the ®rst 100 generations.............. The donation rate is the percentage of pairings in which an agent cooperates by making a donation.....2 79.... ensuring the formation of a dominant cluster with similar tags... We can measure the relatedness of a dominant cluster by the proportion that has its modal tag.2 Average tolerance 0... The donation rate and tolerance still decrease with cost......... Thus there continues to be a pressure towards donating only to those with quite similar tags.. We note in Table 1 that when there are more than two pairings.... followed by a recovery when a more discriminating individual succeeds... compared to only 2% shown in Table 1....... the relatedness of a cluster when it ®rst becomes dominant averages 79%..... The success of tag-based donation requires enough pairings per generation to establish dominant clusters.. The vulnerability of the dominant cluster is realized when a mutant's tag happens to be within the range of tolerance of the typical member of the dominant cluster..018 0.......... the donation rate jumps to 74%.. This makes possible the overall donation rate of 73.. the members of the dominant cluster are vulnerable to a relatively intolerant mutant....... Higher numbers of pairings also increases the chances that similar agents will continue to ®nd agents to donate to. In all these simulations the typical behaviour of the system is attained within a few hundred generations..... With this method.. Tags are initially chosen at random.....nature..... The cycle continues inde®nitely.. The sharp transition suggests that it may be possible to approximate these simulation patterns in an analytic model......... the Table 1 Effect of pairings on donation rate Pairings 1 2 3 4 6 8 10 ...... and adopts the other's tag and tolerance if the other's score is higher than its own.......... the average tolerance of its members tends to drift upwards.... even one pairing is suf®cient to achieve a donation rate of 49%.. with a similar tag.... and soon establishes a new cluster of agents with similar tags and similar (low) tolerances...3 73.. The average donation rate recovers quite quickly........000 generation period history.... As the members of this new dominant cluster do not contribute to the old cluster. Once this happens.......... they all inherit the same tag.....019 0.. the average level of tolerance increases..5 60......... the old cluster dies out and the average donation rate rebounds............. unconditional cooperators................ donates b 1:0 at a cost of c 0:1 to itself....... the donation rate falls when the relative cost of making donations rises........007 0... Beyond a cost of 0......... For costs less than 0. the recipient gets b 1... 442 © 2001 Macmillan Magazines Ltd NATURE | VOL 414 | 22 NOVEMBER 2001 | www..... The dynamics do not end with the establishment of a dominant cluster of similar agents who help each other. on the donation rate (the amount of cooperation) and on the population's average tolerance T........ This establishes the evolution of cooperation without reciprocity..... The result is that the fortunate mutant has many offspring over the next few generations.... The members of the new cluster donate to each other when they happen to interact because.........7 73.. P....... Average tolerance is calculated over the entire population..... that is. Our model of donation based on similarity of tags extends the insight of Nowak and Sigmund10 by reducing the requirements for the participating organisms: a potential donor incurs a certain cost in order to help another individual if and only if their tags are within the donor's range of tolerance............... As agents participate in more pairings.. Cooperation based on tag similarity does not require that the agents Table 2 Effect of cost of donating on donation rate Cost 0.....com .019 0...........................6 76...... In fact...letters to nature spring begin to spread through the population...... cycle of increasing and decreasing tolerance could re¯ect... The cost/bene®t ratio........7 2..... This fortunate but relatively intolerant mutant will then tend to score very well by receiving help from most of those who pair with it.......................2 Average tolerance 0..
. Econ... R.. Science 211....0. recovered Davis Cambridge Proportion recovered Davis Cambridge Recovery accuracy* Davis Cambridge No.19 6 1...07 6 0. M.... Oxford.... Indeed... 1995). 349±350 (2000)..... but only if they had been watched during the caching trial. Lond. & Hamilton..56 6 0. A....46 3..0.32 0. Soc..). Science and the Arts Enrichment Fund. D. Massachusetts..69 1.... P. J... Social Networks 11. New Mexico. To test this hypothesis.e. R.. 783±784 (2001).. W.06 0. 16.20 2. Food storing may reduce this competition as the food can be collected quickly and hidden elsewhere2±4.. The cultural evolution of bene®cent norms. Genetic Algorithms (ICGA97) (ed. The emergence of social organization in the prisoner's dilemma: how context preservation and other factors promote cooperation... 278±307 (1996). even though they had observed other jays caching. 25.. R. Rationality Soc. The evolution of indirect reciprocity.. Soc..07 0.28 6 0..21 6 1.. D. Genetic imprintingÐurinary odour preferences in mice.0...43 4. Strategies of donating to others who have suf®ciently similar heritable tagsÐeven though such tags are initially arbitraryÐcan establish cooperation without reciprocity..21 6 0....... our results show that cooperation can become established and be sustained even without memory.. cached Davis Cambridge No. In the wild. Linster....... Self-referent phenotype matching in a brood-parasite: the armpit effect in brown-headed cowbirds (Molothrus ater).. 5.. such as for chemical markers or cultural attributes. however.. 7..C.. Keller.0. New York. Dawkins.... R... Evolutionary biologyÐgreen beard as death warrant... J.09 3...5 . & Sigmund.. 4. & Paprika....... M.08 6 0. Like other corvids6±9.0....0. 695±700 (2000).. kin recognition may use tag-based mechanisms such as the `green beard'1.......... Grafen.... H. Sherman. R... University of Cambridge. 23. Trivers..... and then recover their caches in private. Reading. Cogn... S. & Ross.... Biol. G. R....... and speci®cally in new sites unbeknown to an observer.. We therefore predicted that they would be more likely to re-cache any uneaten food...48 6 3.... & Sherman.. 7....20 0. & Johnston. 826±829 (1992)....05 . Holland..... E.. D. Reading.44 6 0. Hamilton.. 573±575 (1998)...0.. For computing facilities we thank the University of Michigan Center for the Study of Complex Systems.. Proc..... jays were allowed to cache either in private (when the other bird's view was obscured) or while a conspeci®c was watching. Evol. Natl Acad. Therefore cooperation on the basis of similarity could be widely applicable in situations where repeated interactions are rare... Correspondence and requests for materials should be addressed to R.51 0. it is a risky strategy because caches can be pilfered by others5±9. 11.. M.S..20±23 and `armpit' effects24±28...m.. The Evolution of Cooperation (Basic Books. Nucleus and shield: the evolution of social structure in the iterated prisoner's dilemma. Scrub jays (Aphelocoma coerulescens) remember `what'.. 6547±6551 (1996). Biol.... Nature 394.52 6 0..... Cambridge CB2 3EB........ Nature 394...... R. L.. The evolution of cooperation. Ef®ciency in evolutionary games: Darwin. Axelrod. Cohen. L. 58... G... scrub jays were allowed to cache wax worms in a sand-®lled caching tray during two Table 1 Behaviour of the observer + pilferer birds during observed and in private caching treatments Caching treatment Behaviour Observed In private n Wilcoxon pairs test Z P . Nature 359.1 ...01 1. We proposed that birds re-cache to minimize potential pilfering by observers.... they do not even need to observe the behaviour of others or receive reports from third parties..... 61.. The basis for similarity also can be `secret handshakes' or other arbitrary behavioural signals that individuals can detect19. Indeed. Ma.. 18.93 5.... Kin recognition and the `armpit effect': evidence of self-referent phenotype matching. Our results suggest that jays relate information about their previous experience as a pilferer to the possibility of future stealing by another bird.07 4. M. 1991).) 295±312 (Addison-Wesley.. 8... re-cached Davis Cambridge Proportion re-cached Davis Cambridge 8.05 . USA 93.. Effects of experience and social context on prospective caching strategies by scrub jays N. Hauber... (e-mail: rlriolo@umich. A... L. Ba Kaufmann..... M... & Axelrod.05 ..0. Mateo.08 2. 521±523 (1998).. 13..0.. Riolo.20 2..5 .. 113±117 (2000). Oscillations in the evolution of reciprocity. 28.. The Sel®sh Gene 96 (Oxford Univ.....68 2...... Biol.03 0. Social life has costs associated with competition for resources such as food1...... in Arti®cial Life II (eds Langton.74 6 1.71 6 0.. In summary... Trends Ecol....24 0.. The genetical evolution of social behaviour. Treatments consisted of three trials/caching treatment at Davis. 1976)... J.....05 ..... J. P. New York. Clayton² * Sub-department of Animal Behaviour. A Wilcoxon matched-pairs test compared the effect of caching treatment for each of the behaviours listed. R.. and readily re-cache them in new places when the observers are no longer present (for example ravens8..com No. R.. The Evolution of Cooperation 146±150 (Basic Books.... M Received 3 August... Cambridge CB3 8AA.. 1984). W. Q... 1±52 (1964)..... R.. Nowak. 279±301 (1992). K... Press.. W.. Sci.. 213±236 (1989)... R. R.. Working paper 99-01-002 (Santa Fe Institute.letters to nature are able to recognize each other from past interactions. South. 15.... J.... K... Cultural artefacts that can serve as tags include accents. food-storing corvids return to caches that they had hidden in the presence of conspeci®cs... UK ² Department of Experimental Psychology.... in Proc. E. 12...70 6 0. followed by three trials/caching treatment at Cambridge (data are mean 6 s.. To test this.9.... Nor does it require that one agent can observe and recall how the other agents behaved with third parties....... 21±26 (1989).......20 ... unpublished observations).... Nature 393.08 2..... 21..10 6 3.84 4....0.10 0..61 6 0. D.. Riolo. M....... Nowak. 1997). Axelrod.. Baum.. T.56 0...55 10.... Social Forces 71. pilfering them when given the opportunity.0.. & Axelrod..0.....04 7 7 7 7 7 7 7 7 7 7 7 7 2..L.. Isles.38 6 1..01 0. As an agent does not have to remember previous interactions with another agent.) 378±385 (Morgan Riolo.. Alexander. accepted 17 September 2001.... Evolutionary stability in the in®nitely repeated prisoner's dilemma played by two-state Moore machines...20 2.15 1... The Extended Phenotype 146±151 (Freeman... 19. 27. 3. `where' and `when' they cached10±13....05 . Jays without pilfering experience did not. 35±57 (1971).. Evolution of indirect reciprocity by image scoring. Proc...19 0.. W... 13. Cohen. Nowak.. Lindgren.05 3... 22. Biol. M. & Sigmund. K... Nature 409... The evolution of reciprocal altruism. N..95 6 0. and modify their caching strategy accordingly.. 880±903 (1992).. I and II....36 6 0.. 6. R. 9.... and reputations are not established.. Anim. R. R. D.1 . A.. D...19 6 0.03 1.... Allison. Lomborg. Theor.. Emery* & N... E... D.. the basis for similarity can be completely arbitrary........37 0. B.... UK The authors contributed equally to this work .. 379±396 (1990). Gestational drive and the green-bearded placenta.... Here we show that jays with prior experience of pilfering another bird's caches subsequently re-cached food in new cache sites during recovery trials.. Massachusetts.. C. M.... 14. Nash and the secret handshake. D..R. Hidden Order: How Adaptation Builds Complexity (Addison Wesley... P. N.... 1999)...05 ................ Rev. San Francisco. R. R... Robson... Theor...... Rev.81 9.. 7th Int....... Conf.. L..... È ck. B.... Acknowledgements For ®nancial support we thank the Intel Corporation and the University of Michigan College of Literature.....57 6 0. 20. .. et al. European jays15.. Keverne.. D.. & May.. 2...85 0..06 6 0......... Sel®sh genes: a green beard in the red ®re ant..... they remember where conspeci®cs have cached.05 ... The armpit effect in hamster kin recognition.. 17...... K.. © 2001 Macmillan Magazines Ltd 443 . Theor... 1390±1396 (1981).. * Number of looks to ®nd ®rst cache. 137.. 144.. E. Am.. J.... M... A.. an agent only needs very limited signal-detection capability.. B. Not only do the agents not require continuing interactions..nature..edu).. 15. but only when they had been observed caching. San Francisco.... 1982). 1984)... 5±32 (2001)... The Biology of Moral Systems (Aldine de Gruyter.. 1.... The role of social structure in the maintenance of cooperative regimes...71 6 0. J. 24..92 2. 26.. A. 10..57 6 0... scrub jays. Hauber. & Richerson. practices or artefacts subject to fashion such as wearing hats of particular colours18. 46.. B 267........ Haig. P. J............ NATURE | VOL 414 | 22 NOVEMBER 2001 | www........ 1987)... M. Dawkins...19 6 1..... New York. D. University of Cambridge. Evolutionary games and spatial chaos.. J. A.14.. Using tags may also be interpreted as imposing an abstract topology on the agents in which an agent's `neighbourhood' is de®ned by its tag and threshold of similarity tolerance14.1 .. 573±577 (1998).....36 2. J.. but may also adjust their own caching strategies to minimize potential pilfering. & Allen. 3. let alone know anything about that agent's behaviour with others.. Axelrod.. Boyd.
Mon. Astrophys. Cost and benefit are measured in reproductive success. The basic outcome of the computer simulations3 is that a substantial degree of cooperation is established. hiding one’s accent in a foreign language is nearly impossible. UK. D. and then sharply declines. nding them. 252. J. M. B. L5–L9 (2000). C. by new individuals whose traits lie in the range of the dominant cluster but whose tolerance is considerably reduced. Wu. But how does such behaviour evolve? Some new computer simulations provide a plausible answer. R. Astrophys. and in its wake a reduction in overall cooperation. 425 -427 (2001).theevolutionarydevelopmentof a fictitious population of ‘agents’ is simulated. Astrophys. 8. Consider a gene that confers on its bearer not only a green beard (or any other distinctive label).bham. they spread. Because ‘selfish’ individuals. N. could be a reliable tag that is hard to fake. serves to reheat the clusters. whereas Chandra has the spatial 2. But if individuals can guess whether recipients are likely to give assistance in their turn. Gas cooling onto these galaxies serve toof fuel the and black galaxies atthat the this.com . L. Each individual has a trait (or ‘tag’. Hanany. A wave of intolerance then sweeps through the population.ac. Valageas. 383. B. sensitive enough to 7. provided by the new generation of X-ray Trevor Ponman is at the School of Physics and 6. J. cooperation resumes at its former level and tolerance starts spreading again. F. 7. In traditional models of how cooperation canemerge6. E. 8. 319. call it). 104 -111 (1991). Soc. Cannon. it slowly increases over time. Nature 404. Solutions to these problems may soon be provided by the new these harbour central black holes. L. but with some crucial differences. Not. consisting of players sufficiently similar to help each other. Astron.. © 2001 Macmillan Magazines Ltd NATURE | VOL 414 | 22 NOVEMBER 2001 | www. J. 135–137 (1999). simultaneously. ■ 414–441 (1991). a signal and a predilection for the signaller. Edge. Astrophys. D. J. Individuals with such a gene would effectively form a self-serving clique.Wu. 318. and getting help provides a larger benefit to the recipient. ing onto these galaxies serve to fuel enough to measure themay amount of gas inthe galaxy groups. hole. like a dress code). maximum tolerance means helping everyone. But it still takes some effort to accept the idea of a gene producing. and so the gene would spread within the population. T. the green beard is a cherished icon of the ‘selfish gene’ view of natural selection. than supernova explosions. et al. in some way.J. the mechanism that leads to cooperation is a form of kin selection — either classical (if traits are inherited genetically) or social (if they are inherited culturally. J.sr. In other words. C. a ‘dominant cluster’ emerges. & Henry.3 discuss a more plausible model for the evolution of cooperation: W individuals just need to like their like. & Nulsen. This drop occurs when the dominant cluster is dissolved from within as a result of mutation. We are witnessing a wave of social and religious intolerance right now. Bright Trevor Ponman may is at the School Physics harbour central black holes. Astrophys. S. This Thisis isnot not easy. e-mail: tjp@star. Fabian. Such a double-effect gene seems contrived. C. based on reciprocation. with pairs of individuals meeting randomly as potential givers and receivers of help. Evrard. Ponman. active galaxies and the gas surrounding 2. 955–959 (2000). 425–427 (2001). G. 933–938 (2000). however.E. G. Nature 397. A.now Gas coole-mail: generation of X-ray observatories in operation intjp@star. an ability that is widespread among animals7 (odours or visual cues can provide the required information). & Ponman. 9. E. & Stewart. 1. It will be important in the future to explore the robustness of this phenomenon. E. J. Evrard. These oscillations of tolerance levels are striking. Mon. L5 -L9 (2000). Astronomy. who refrain from helping. 545. centre of many Astronomy. Astron. Nature 414. A. Astrophys.P. R. just like Axelrod’s famous computer tournaments based on the ‘prisoners’ dilemma’ game8. de Bernadis. The new scenario applies to both genetic and cultural tags9. J. 4. to reduce the complexities of political life to the vagaries of a virtual population. 10. J. Kaiser. G.10 the low-entropy gas . et al. et al. incur no costs. retical developments over the next few years. Astron. so they bear fewer costs than the established residents. without direct or indirect reciprocation. This is not the first time that the idea of ‘like helping like’ has been suggested as a route to the evolution of cooperation. P. Not. over many generations. to study the interaction between central 10.news and views them. T.. be expected to stimulate further 9. something that most of us can relate to. Fabian.bham. and is endowed with a tolerance level. These observational advances can 4. 955 -959 (2000). University of Birmingham. These observational can be galaxy groups. 350. 319. 350. Birmingham B15 2TT. they can channel help towards those who help. & Bryan. J.nature. Riolo et al. black hole. Dawkins4 introduced the ‘green beard effect’ as a thought experiment in sociobiology. H. Soc. tribal customs or religious creeds are all tags that induce cooperation. & Stewart. & Bryan. 383. Ponman. R. Part of its appeal is its obvious link to reality — school ties. have a tendency to cooperate and help each other.sr. S. easy. 5.A. with some mutations occasionally introducing new variations into the population. 933 -938 (2000). Cyclically. So. and it isthe possible that this. S. J. Language. Not.. 889 -912 (2000). Some of these tags are easy to fake and might invite exploitation. 95 -103 (1991). Riolo and colleagues3 discuss a new model for the evolution of cooperation. University of Birmingham. Nowak Humans. 414 -441 (1991). Voit. 383. measure the amount of gas in galaxy groups. K. and may well lead to a cottage industry of follow-up investigations. 135 -137 (1999). & Navarro. but also the instinct to provide assistance to all other owners of a green beard. in which individuals help others that are. These newcomers are helped by all the residents of the established cluster but themselves help just a few. serves to reheat Edge. G. This occurs even in the absence of repeated interactions and reputation effects — that is. as Riolo et al. S. 383. 889–912 (2000). P. & Silk. Voit.uk space. G. All that is needed is some recognition of what is ‘similar’. can emerge6. et al. J. Mon. R. Hanany. and bring to mind many historical instances.’s model3 follows a similar pattern. 725–742 whereas Chandra has the spatial resolution 9.theoSoc. K. & Ponman. T. C. J. anticipate a solution that was provided by William Hamilton2 more than a century later: cooperation can emerge as a result of ‘kin selection’ in cases where interacting individuals are genetically related. rather central resolution to study interaction between active galaxies and the gas surrou3. and many other species. Mon. Astron. M. Tides of tolerance Karl Sigmund and Martin A. Not. R.is reheating from starof formation may not be whole story. hen Charles Darwin1 published his theory of evolution in 1859. and it is now widely believed that Birmingham B15 2TT. Not. K. Twenty-five years ago.uk P. Helsdon. and offspring are supposed to inherit the parent’s behaviour. reheating fromof star formagalaxies found at the centre many clusters. C. Bright developments over the next few years.&Henry. 252. Cannon. P. Astron. Second. The slow upward drift of tolerance seems to be due to a combination of mutational pressure and kin selection. He did. & Nulsen. 545. S. of course. F. It would be foolish. Astrophys. Astron. A. tags can help to encourage the usual suspects behind cooperation among unrelated individuals: direct and indirect reciprocation (whereby recipients 403 3. and similar effects have actually been found in nature5. Essentially. because galaxy becausetheir their X-ray emission is weaker thanadvances that of large to the stimulate further theoretical X-ray emission weaker than that large expected clusters. Mon. Mon. and it is now widely believed that these tion may not be the whole story. In this way cooperation. ■ clusters. 104–111 (1991). 725 -742 (1999). Astrophys. XMM-Newton is sensitive 1. On page 441 of this issue. J. XMM-Newton is UK. he knew that cooperation and altruistic behaviour present something of a problem for a concept that is based on competition and the struggle for existence. unless mutations occur. observatories now in operation in space.ac. K. Nature 414. J.10 . R. on the other hand. Valageas. Soc. 318. Both the trait and the tolerance level are inherited by the offspring. Donors refuse to offer help if the trait of the recipient differs from their own by more than the donor’s tolerance level. Soc. Helsdon. (1999). and discriminate against exploiters. zero tolerance means that an individual helps only those who are exactly like it. and are it is found possible rather than supernova explosions. Astron. G. like themselves. P. C. Not. A. J. P. Today. This tolerance does not freeze at some fixed value. lowentropy gas9. & Navarro. Astron. club memberships. Giving help entails some cost to the donor. S. Evolution 6. groups. & Silk. 95–103 (1991). Nature 397. T. P. Nature 404. J. and after some generations cooperation will be eliminated. Furthermore. One attractive feature of the new simulations is the evolution of tolerance — the recognition mechanism that discriminates ‘us’ from ‘them’. But once a new dominant cluster is established. Riolo et al. Soc. A. Yet these computer simulations do capture the imagination. on a computer. de Bernadis. N. Solutions to these problems may soon be 5.. J. Kaiser. H.are Second.
The clever part is to work at temperatures low enough for the intermediates to be frozen in time and to use electron injection rather than thermal excitations to make the reaction proceed. L. 98–103 (1995). & Sigmund.. The catalytic oxidation of carbon monoxide (CO) is one of the simplest catalytic reactions. K.6. The formation of the © 2001 Macmillan Magazines Ltd e- STM tip O C O O CO+* O2+* c Reaction CO* O2* O2*+* d O Desorption C O 2O* eO CO O O CO*+ O* CO2*+* CO2* CO2+* Figure 1 Surface chemistry in action. 1997). London. & Axelrod. and might even promote longterm pairings based on similarity10. New Jersey 08540. Nørskov Modern microscopes are not just for imaging. But tag-based intolerance could turn discrimination away from the ‘bad guys’ and raise senseless antagonism. 272. D. 1). 1976). 10. moving the tip into position over an adsorbed oxygen atom. 4. M. can be viewed directly. At these low temperatures the authors use O–CO–O complex is confirmed by observing a change in the vibrational frequency of the CO. making it hard for tolerance to cross the divide. 1c. Sci. Culture and the Evolutionary Process (Univ. The Selfish Gene (Oxford Univ. 1. Riolo. An STM works by applying a voltage to a sharp metallic ‘tip’. We know that discrimination is needed to sustain cooperation in the face of exploiters. 573–575 (1998).ac. In this context. CO is transformed into carbon dioxide (CO2) by reacting with oxygen or nitrous oxide (NO) on the surface of a platinum. and so cannot be observed during a thermal reaction. the flow of electrons is kept constant by adjusting the distance between the tip and the surface. either to the donor or to a third party). The Evolution of Cooperation (Basic Books. e-mail: karl. Nowak is at the Institute for Advanced Study.nature. atom by atom. Darwin. (An asterisk denotes a site on the surface that can form a bond to an adsorbed atom or molecule. e-mail: nowak@ias. Am. P. 9. On the Origin of Species by Means of Natural Selection (John Murray. R. they image the intermediates on the surface — they observe.) 405 NATURE | VOL 414 | 22 NOVEMBER 2001 | www. 1a) can occur on a silver surface. Boyd. R. ■ Surface chemistry Karl Sigmund is at the Institut für Mathematik. New York. and the International Institute for Applied Systems Analysis. New York. and is an important part of the reactions that take place in the catalytic converter in your car. to move individual CO molecules close to the oxygen atoms on the surface. 3. The formation of carbon dioxide (CO2) by catalytic oxidation of carbon monoxide (CO) on a metal surface is an important reaction in air purification. in Proceedings of the 7th International Conference on Genetic Algorithms (ed. M. for instance. Hahn and Ho show that CO and oxygen atoms (O) adsorbed on a silver surface can be transformed into CO2 with the help of an STM tip. A-2361 Austria. The surface acts by adsorbing the reactants and encouraging them to react until the products leave the surface. In the right hands they can be used to follow and control catalytic reactions on a metal surface — one atom at a time. but it is analogous to the way natural catalysts (enzymes) manage a Adsorption b Dissociation of O2 O O O C S a scanning tunnelling microscope (STM) to make the reaction happen and to image and manipulate the atoms and molecules on the surface (Fig. and applying a new voltage pulse with the opposite sign to transfer the molecule back to the surface and to kick start the reaction. palladium or platinum–rhenium catalyst3. The use of powerful tools is now helping to change that. W. Strudlhofgasse 4. But tags can also present major obstacles in overcoming segregation. any territorial distribution would favour such ‘speciation’. The STM can also be used to physically modify the surface: the electrons transferred from or to the tip can start a chemical reaction. R. R. W.news and views of help return the help. Keller. The reaction involves five steps. Nature 393. 7.at Martin A. 7. many aspects of catalysis by solid surfaces are not understood. These experiments provide atomic-scale details of a chemical reaction occurring at a surface. 2. K. olid surfaces act as catalysts for a large number of reactions: it is estimated that 20–30% of the gross national product in developed countries is dependent one way or another on this sort of catalysis1. Catalysis frozen in time J. Biol. C. With this technique the reaction can proceed at temperatures so low that every step of the reaction (a–d) can be watched closely. Hahn and Ho2 show that they can manipulate individual atoms and molecules adsorbed on a metal surface to induce a catalytic reaction. In practice. Pushing or pulling an individual atom or molecule with the tip can also make it move. Cohen. Indeed. the two first steps (Fig. d). Theor. In other words. 573–577 (1998). and then recording the flow of electrons that tunnel between the surface and the tip. to find out how much one has in common is one of the first delights of falling in love and contemplating a lifelong partnership. 6. and to induce the final reaction and desorption of the product (Fig. Yet despite its importance. A. Measuring the spatial distribution of the vibrational intensity of CO within the complex provides information about the structure of the reactants. Axelrod. or make atoms or molecules move on the surface5. The work of Hahn and Ho shows that it is possible to induce reactions that won’t proceed thermally by using tunnelling electrons to activate the reaction. 1984). & Ross. 1b).) 378–385 (Morgan Kaufmann. T. 5. A-1090 Vienna. They also used the STM in its ‘inelastic electron tunnelling spectroscopy’ mode to monitor the vibrational behaviour of CO as it is nudged closer and closer to the two oxygen atoms.sigmund@univie. Nowak. Riolo. Universität Wien. K. D. but the rest of the reaction requires higher temperatures to proceed spontaneously. Nature 394. which is scanned across the surface. Laxenburg. W. San Francisco. Dawkins. R. At the low temperatures (13–45 K) used by Hahn and Ho. In a separate set of experiments they induced the reaction by first transferring the CO molecule to the tip (by using a voltage pulse). & Sherman. Using an STM tip to induce chemical reactions at a catalytic surface is not an efficient way of producing large amounts of chemicals. USA. P. a record of its height at each point provides a detailed map of the surface at atomic resolution4. Princeton. Hahn and Ho use the STM to dissociate the oxygen molecules (Fig. emission control and chemical sensing.3 do not produce dominant clusters that split into rival tribes. L. Bäck. Although the simulations by Riolo et al.com . tags bolster the emergence of cooperation in repeated interactions. which are summarized in Fig. Nature 414. Along the way. Press. 1–16 (1964). By using this technique at low temperatures they can control the speed of the reaction. Pfennig. which transfers electrons to the reactants. When the tip is scanned over the entire surface. 8. Tags would then act as selfenforcing stereotypes. R.edu 1. a complex consisting of two oxygen atoms close to each other and to a CO molecule. Hamilton. Chicago Press. so they can follow important steps as they happen. 1985). Intermediates that would not have a measurable lifetime at higher temperatures. & Richerson. 1859). In their experiment2. 441–443 (2001). J. D. Writing in Physical Review Letters.
is thought to be the key to establishing cooperation among non-relatives.nature. and show that cooperation tends to collapse when individuals bearing identical tags are given the R NATURE | VOL 418 | 1 AUGUST 2002 | www.4 have presented a model in which cooperation is instead based on similarity: agents donate only when their partner’s ‘tag’ lies within a ‘tolerance’ range around their own. But Riolo et al.com/nature © 2002 Nature Publishing Group 499 . whether direct2 or indirect3.brief communications Behavioural evolution Does similarity breed cooperation? eciprocity1. Here we point out that their model requires individuals with identical tags to cooperate with each other.
Theor. under some conditions. 2. the level of cooperation for other parameter settings is substantial. 4.000 generations. Canada 1. We therefore question their mechanism for maintaining cooperation without reciprocity. although. J. Gilbert Roberts*. Nevertheless. Ontario K1S 5B6. Allowing similar individuals not to donate caused cooperation to be restricted in our system. but again the extent of cooperation depends on many factors. USA e-mail: rlriolo@umich. we ensured that all positive mutational changes converted T back to the cooperative region. 441–443 (2001). W. & Nowak. M. If unconditional defection is introduced by adding a binary trait that controls whether agents never donate. 379–396 (1990). including the rate at which ‘never donate’ agents are created. A striking characteristic of Riolo and colleagues’ simulations was the formation. clusters of cooperating agents with similar tags arise intermittently. D.? Once the constraint that identical tags must cooperate has been removed. Cohen†. It was not. agents can have identical tags without having a recent common ancestor. Riolo et al. D. 1976). Biol. even to those with identical tags. For instance. even without any mechanism for cheating through faking tags5. the system of ‘like helping like’ breaks down. the number of pairings. Robson. However. L.m.6%. 3. 2. M. 1 of ref. A. Agents with negative T values would not donate to any other agent. Soc. 5. and to what extent. Whether it does. 1). tag mechanisms are one viable approach.R. Lond. 441–443 (2001). as one tag group is invaded and thus dies off. Nowak. A. 403–405 (2001). Henry Wellcome Building for Neuroecology. J. UK ‡Present address: Department of Biology. Nature 414. Dawkins. which is why..3. through differential reproduction. New York. with the overall rate of cooperation depending on the relative dynamics of invasion. In 30 runs of our modified model. Although they are correct in noting that in our model1 an agent will always donate when it meets another with an identical tag. and the role of signals in such systems will be an important consideration. Because tag similarity is no longer a reliable guide to behaviour. Science 211. UK e-mail: gilbert. Axelrod. R. and ‡Gerald R. the cost/benefit ratio of donations and the particular adaptive mechanisms in the model. The Selfish Gene (Oxford Univ. Q. resulting in very low overall donation rates1. & Axelrod. Riolo. M.2. Our results show that. might be difficult to copy3. Cohen. However. so in our modified system they can share tags without sharing the rule for cooperating. We found that introducing the realistic option of non-donation had a catastrophic effect on cooperation (Fig. Whereas the problematic ‘green beard’ effect7 depends on a link between altruism and a particular trait. We believe that the difference has not been fully understood between the stability of cooperation within any particular tag group and the rate of cooperation across a population consisting of diverse tags with changing frequencies over time. mutants that fail to donate. Nature 393.nature. Rick L. 573–577 (1998). 6. depends on a link between altruism and similarity. D. All parameter values were the same as for Fig.roberts@ncl. as in their version of the model. whereas other tags. What makes cooperation so challenging for theorists is explaining how it can persist in the face of more exploitative strategies. & Hamilton. However. in which individuals with identical tags must donate (blue). 35–57 (1971). but they can do even better by accepting donations without donating. under some parameter settings (few pairings or high cost of donation). such that they accept more donations than they offer. similarity can indeed breed cooperation. Sigmund. Michigan 48109. another tag group with more reliable cooperators can flourish and become dominant. University of Durham. R. W. 0.). M. so a form of kin selection6 can support cooperation. we do not believe that their basic claim is correct. resulting in the cycles of cooperation and tag dominance noted previously1. tags that are easy to copy might lead to high rates of invasion. reply — Roberts and Sherratt argue that if agents with identical tags are allowed a choice of behaviour. South Road. R. L. Ottawa. we replicated Riolo and colleagues’ simulations with one simple modification: we allowed tolerance to evolve to below zero. only to be undermined by agents that reduce their tolerance level. K. Rev. depends on several factors. Nature 414. For example. This means that when individuals with identical tags interact. territorial distribution of agents might favour ‘speciation’ into selfenforcing stereotypes3. Proc. 46. R. University of Newcastle upon Tyne.48% (s. Sigmund.L. G. We find that if mutations are not biased as strongly © 2002 Nature Publishing Group towards ‘never donate’. School of Biological Sciences. To investigate what would happen if agents were given the option of declining to donate to any other agents. Nevertheless. 427–431 (1998). 1390–1396 (1981).edu 1. R. 7. Cooperation under the original conditions of Riolo et al. D. 7. T. they must always donate.com/nature .. operates through a process of ‘like helping like’5 — agents sharing any particular tag also share the rule of donating to each other.ac. M. we believe that possible mechanisms by which cooperation can arise without reciprocity merit further attention8. each for 30. Cooperation based on similarity there500 100 Donation rate (%) 75 50 25 0 0 100 200 300 Generations 400 500 Figure 1 Population dynamics for the first 500 generations of a typical run of Riolo and colleagues’ model4. Thus. 403–405 (2001). A. 3. & Sigmund. 1–52 (1964). K. Newcastle upon Tyne NE2 4HH. Ann Arbor. Trivers. B 265. Theor. Cohen. In the system of Riolo et al. †School of Information. J. and our modified model in which individuals with identical tags may or may not donate (red). Biol. R. Many factors could affect the dynamics generated by tag-based mechanisms. Nature 414. cooperative periods are rare and short-lived.’s 73. 8. 4. Hamilton. of clusters of agents with identical tags. Most individuals in their simulations (up to 97% of the population4) were therefore ultimately constrained to cooperate.uk †School of Biological and Biomedical Sciences. Our model could also be extended to study how a tag mechanism acts in conjunction with other mechanisms known to affect the emergence of cooperation. A. Why do we not find the high degree of cooperation reported by Riolo et al. Michael D. & Axelrod. K. the overall mean donation rate was 1.. even those with identical tags. Durham DH1 3LE. University of Michigan. in comparison with Riolo et al.031%). such as language or accent. then tag similarity can no longer be a reliable guide to behaviour and so similarity does not breed cooperation. R. the system of Riolo et al. Sherratt†‡ *Evolution and Behaviour Research Group.e. We have replicated the results of Roberts and Sherratt and have run a generalized model that includes theirs as one extreme and our original model as another (details are available from R. or donate using tags and tolerance. Riolo*. we find that cooperation also emerges. 144. there is a limit to such cheating imposed by the minimum Tǃ0. Press. by setting the minimum boundary for T at ǁ10ǁ6. Biol. Roberts and Sherratt claim that cooperation based on similarity was built into our model. Ford School of Public Policy.brief communications option of not donating. Nature 414. destroying cooperation. There is no dispute that particular cooperative tag groups are invadable1. agents interacting with others bearing the same tag face the classical ‘prisoner’s dilemma’ — they can do well by cooperating. Thomas N. Riolo. R. NATURE | VOL 418 | 1 AUGUST 2002 | www. Carleton University. will tend to invade. & Nowak. fore turns out to be a rule that was built into the model rather than an inference that can be drawn from it. Roberts. resistance and emergence of dominant groups. Robert Axelrod‡ *Center for the Study of Complex Systems. Further investigation is needed to understand fully the range of mechanisms that can produce cooperation without reciprocity.
This action might not be possible to undo. Are you sure you want to continue?