The Evolution of Cooperation

Darwinian evolution proceeds through natural selection, whereby organisms whose inherited characteristics are most conducive to survival and reproduction will – for that very reason – tend to leave more descendants than others, so that those inherited characteristics will increase in relative frequency within the population from generation to generation. This logic seems to imply that it is individual reproductive benefit that drives evolution, not benefit to the population or group.

An Illustration of "Selfish" Evolution

Suppose, for example, that a population of animals – let's call them "dingbats" – inhabits a small isolated island with limited resources. The population can thrive as long as it maintains a steady level, but let us imagine that a particular dingbat is then born with a genetic mutation B which leads it to breed at twice the usual rate (for the thought-experiment it does not matter how or whether this could plausibly happen). Barring accidents, this dingbat will leave more offspring than the average, and each such offspring has a 50% chance of inheriting mutation B. The same will apply to those offspring, so from each generation to the next, the relative frequency of B within the population will tend to grow, implying an increased rate of reproduction overall. Since resources are limited, there will be increasing competitive pressure for survival, and so the continuing relative increase in B-dingbats over time implies an absolute decline in non-B dingbat numbers, and they will ultimately die out. At the same time, it might well be that the resources of the island prove unable to sustain the new speed of dingbat reproduction, so that edible plants die because their shoots are being eaten too soon, or animals on which the dingbats prey are all killed off. So in the end, the entire dingbat population dies out, having exhausted its environment.

In this thought-experiment, it is clear that short-term "selfish" genetic benefit to the B-dingbats (i.e. their tendency to out-compete the rest of the dingbat population) dominates any long-term benefit to the population as a whole. Natural selection operates through immediate competition for survival and reproduction, and cannot look ahead to what will ultimately be best for the population (or even for the B-dingbats themselves). Evolution does not – as commonly presumed – automatically lead to "the improvement of the species". This recognition is what lies behind the general rejection of "group selection" by George Williams, William Hamilton and Richard Dawkins in the 1960s and 1970s. And it is obviously tempting to draw analogies with human society: the fact that one national or social group out-competes another does not in any way imply that it is "superior" in any respect other than short-term domination. Huge military spending or rapacious economic exploitation of the world's resources might achieve such domination over others, but nothing guarantees that this will lead to a situation of long-term benefit for the human race (or even for the successors of the dominant group).

Is the Evolution of Altruism Impossible?

These rather depressing reflections might naturally seem to imply that genuine altruism – behaviour which leads to individual sacrifice for others' benefit – is almost an evolutionary impossibility, at least beyond the immediate family. A tendency of individuals to sacrifice their own well-being for that of their offspring or close relatives (so-called kin selection) is easily explained: a gene R which led to such behaviour could easily spread within a population, because offspring carrying gene R would tend to be favoured by their parents' possession of gene R (and consequent relative-favouring behaviour). But humans and other species (arguably vampire bats, for example) seem to exhibit altruism even to unrelated individuals. Is this some sort of illusion – really a form of disguised selfishness – or might there be some way in which real altruism could have evolved after all?

The Prisoner's Dilemma

Evolutionary game theory investigates the evolution of behaviour in terms of games and the consequent "payoff" to each party, and has become a major focus of evolutionary biology. Of the various games in the literature, by far the most famous is The Prisoner's Dilemma. Stripped to its essentials, this involves two individuals, who are unable to communicate with each other, but each of whom has to decide whether to "cooperate" with the other (i.e. act favourably towards them) or "defect". Depending on what they choose, they get a "payoff", which is presented in the table below from a first-person perspective:

He cooperates He defects
I cooperate 3 (Reward) 0 (Sucker's Payoff)
I defect 5 (Temptation) 1 (Punishment)

In this situation, it is clear that whether the other person cooperates or defects (and I cannot know in advance which he will have chosen), I am better off defecting. Since the setup is symmetrical, he is in the very same position with respect to me. So it seems that if we are both "rational" self-interested players, we will both end up defecting. However we then both get a payoff of only 1, whereas if we had both "naively" cooperated, we would have both been better off, with a payoff of 3. This seems paradoxical, in that rationally self-interested behaviour leads, for both of us, to a worse outcome than if we had "irrationally" decided to be altruistic.

The Iterated Prisoner's Dilemma

If the relevant context involved nothing but unrelated "one shot" Prisoner's Dilemmas, then it seems that by far the most successful policy – and hence the only policy that could possibly be favoured by evolution – would be uniform defection. But if we consider an ongoing series of Prisoner's Dilemmas (i.e. an Iterated Prisoner's Dilemma or IPD) – in which future behaviour can potentially be conditioned by earlier outcomes, the story is very different. For then we have a wide range of possible strategies, which are not restricted to the simple binary choice between "always cooperate" or "always defect". And it becomes an interesting question which strategies bring most evolutionary benefit – a question which can only plausibly be explored through computational thought-experiments.

The first major experiments of this kind took place through computer tournaments organised by Robert Axelrod, in which he invited various prominent game theorists to submit IPD-playing programs to "compete" against each other. Each of these programs played for 200 iterations with each of the others, and then the overall aggregate scores were compared. The best-performing strategy in both tournaments was the very simple Tit-for-Tat, which started off by cooperating with each partner, and then in subsequent iterations, did whatever its partner had done on the previous iteration.

Tit-for-Tat is a "nice" strategy, in that it is never the first to defect: it will defect only in response to a previous defection against itself. It is also a "provocable" strategy – if defected against, it will "punish" its partner – but at the same time it is "forgiving": if its partner recommences cooperating, it will reward this immediately by reciprocating. It is also a very simple strategy, so that the retaliation and forgiveness are easy to identify and respond to. These all turned out to be characteristic properties of the most successful strategies in Axelrod's tournaments. Most striking was the fact that all of the "nice" strategies tended to outperform all of the "nasty" strategies, a result which completely flew in the face of the widespread expectation that only selfishness could be evolutionarily beneficial. At the very least, this implied that the evolution of altruistic behaviour need not be paradoxical (thus serving as an existence proof of a possible evolutionary mechanism); indeed it strongly suggested that altruistic behaviour ought positively to be expected to evolve, given the appropriate circumstances (e.g. agents repeatedly interacting with each other, and able to respond to each other's previous behaviour).

Evolution and Ethics

All this has obvious implications for the evolution of morality, since it suggests not only a possible evolutionary basis, but also a potential explanation of some of morality's most prominent characteristics (e.g. an emphasis on altruism, retribution, forgiveness, and simple rules). A huge literature has since developed and continues to grow, some of which can be found through the links above. There is also a wide range of relevant computer models available on the web, including the following:

Robert Axelrod

Robert Axelrod

Pioneered modelling of the Iterated Prisoner's Dilemma