快猫短视频

YOUR plane crashes and you find yourself stranded in the middle of a vast jungle. How would you work out which fruits are safe to eat and where to find clean water? You could muddle along on your own for a while, but you would probably end up sick and very hungry. Far better to find some friendly locals and learn how they do things.

Learning from others is something we do all the time, not just in extremis. We are more reliant on so-called 鈥渟ocial learning鈥� than any other animal 鈥� it is thought to be at the core of culture and tradition and is credited with our successful colonisation of the planet. Yet no one knows exactly how social learning works. Obviously, copying others allows us to acquire useful knowledge without having to bear the costs of working everything out for ourselves. But there is a catch. If societies are to adapt to changing conditions, there must be innovation too 鈥� people cannot blindly copy everything because the information may be wrong, outdated or unavailable.

This problem has occupied of the University of St Andrews, UK, for some time. 鈥淚ndividuals ought to be selective with respect to when they rely on social learning and from whom they learn,鈥� he says. 鈥淣atural selection ought to have fashioned specific adaptive learning strategies.鈥� But what are these strategies? If social learning is such a powerful force in our species鈥� success, surely we need to know when, where and why it happens. Yet previous attempts to answer these questions have only scratched the surface. Laland realised that if he was going to get anywhere he would have to come up with an original approach.

Until then, only a tiny fraction of the possible learning strategies had been investigated. The most thoroughly researched was the 鈥渃onformist transmission model鈥� 鈥� the idea that a person is more likely to copy traits that are common in the population than those that are rare. An alternative is 鈥渃opy an expert鈥�, which seems like a reasonable rule to follow when buying a new computer or shares on the stock market, for example. 鈥淐opy the most successful鈥� also makes intuitive sense, although in our celebrity-oriented world there is a chance it might backfire 鈥� George Clooney may endorse a certain brand of coffee, but does he really know any more about beverages than the next person?

Let battle commence

Laland wanted to consider a much broader range of strategies and, crucially, to find out which ones work best. He realised he could not do that with a traditional experiment, so he hit on the idea of holding . His inspiration came from a series of open competitions held in the 1970s to examine why cooperation evolved. These tournaments, based around the prisoner鈥檚 dilemma, which involves deciding when to cooperate and when to defect, were a shot in the arm for research into cooperation. Laland hoped a tournament could be just as successful for social learning. 鈥淲e thought if we were to advertise this idea widely we could attract all kinds of people into the field,鈥� he says. So, teaming up with several other experts in social learning, Laland secured enough funding from the European Union to pay for the project, including a 鈧�10,000 prize for the tournament winner.

Their competition was going to be a game of survival, taking place in a computer-generated world. Virtual agents would have the potential to acquire 100 possible behaviours, each with a different associated pay-off that would change over the course of the game. The pay-off represents the benefit an individual gains by performing a particular behaviour, its changing value reflecting the fact that information can become outdated as the environment changes.

Entrants to the tournament would start with 100 agents each, which would accumulate a repertoire of behaviours over their lifetime through learning. At every round of the game, each agent would have three options: innovation, in which they randomly acquired a new behaviour by individual learning; observation, in which they acquired a new behaviour by social learning; or exploitation, in which they used a previously learned behaviour and so gained its pay-off. The entrants had to devise a strategy that their agents would use to decide between these options. The challenge was to create the strategy that generated the most successful or 鈥渇ittest鈥� agents 鈥� a criterion measured by dividing an agent鈥檚 accumulated pay-off value by the number of rounds it had survived. Furthermore, in each round, every agent would have a 1 in 50 chance of dying. The deceased would then be replaced by an 鈥渙ffspring鈥� of another agent. Agents were chosen to 鈥渞eproduce鈥� with a probability proportional to their mean lifetime pay-off. So the better a strategy鈥檚 performance, the bigger the share of the population its agents were likely to have. By this simulated version of natural selection, the entrant with the most successful strategy would have the most agents at the end of the game.

There were two phases to the tournament. The first was a round robin where all strategies played each other for 10,000 rounds in pairwise contests. The strategy with the most agents at the end was the winner. Then, in the second phase, the 10 highest scoring strategies were thrown in together to see who would win overall. They battled it out in a variety of simulated environments, differing in such parameters as the number of agents a potential learner was able to observe, the likelihood that an agent using social learning would pick up the wrong information, and the way in which pay-offs associated with behaviours changed over time. The aim here was to test how robust the strategies were in different learning environments because in the real world the costs and benefits of social learning versus individual learning may vary.

And the winner is鈥�

The competition, held last year, turned out to be an irresistible challenge to many, with over 100 entries submitted from a variety of academic disciplines, ranging from philosophy to computer science, and even some school pupils. In fact, two teenagers from Westminster School in London beat most of the academics to come tenth overall.

Last month, Laland and colleagues published their findings in Science, . So what did they discover? It seems a successful strategy rests primarily on the amount of social learning involved, with the most successful agents spending almost all their learning time observing rather than innovating. However, avoiding spending too much time learning either socially or individually was just as important. 鈥淏etween a tenth and a fifth of their life seemed to be the optimal range,鈥� says fellow organiser Luke Rendell, also from St Andrews University. 鈥淚f they did more learning than that it seemed that life was just passing them by.鈥�

鈥淭he most successful agents spent almost all their learning time observing rather than innovating鈥�

Successful strategies were also good at spacing out learning throughout the agents鈥� lives. The winning strategy, Discount Machine, submitted by PhD students Daniel Cownden and Timothy Lillicrap from Queen鈥檚 University in Ontario, Canada, stood out because it did just this. It seems packing all your learning into the early part of your life is not a great idea 鈥� we need to keep updating our knowledge as we go along.

Lillicrap points out that the questions their strategy addressed resemble those posed in real life. 鈥淲e face similar trade-offs all the time 鈥� for example, how much education should I get before I join the workforce?鈥� To answer such a question we need to consider various factors such as how much more do I expect to earn with this training? How long is it going to take? What鈥檚 the likelihood that my training will become irrelevant? How long will I be in the workforce? 鈥淥ur strategy takes those things into account,鈥� he says.

Another attribute of the most successful strategies is that they are parasitic. This is the essence of social learning 鈥� somebody has to do the hard graft to find out how to do things before other people can copy them, so it only pays to learn socially when there are some innovators around. Indeed, in contests where Discount Machine agents were able to invade the entire population, they actually ended up with a lower average pay-off than they did in contests where the conditions allowed some agents with more innovative strategies to survive, so providing new behaviours to copy.

This also has real-world implications. Could it be that we don鈥檛 all use the same optimal social learning strategy? 鈥淚t鈥檚 quite clear that you would expect social learning to evolve and be favoured,鈥� says Laland. But if everyone relied heavily on it then there would be a decrease in the population鈥檚 fitness and subsequent advantages for individuals who are more inclined to learn for themselves.

General observation certainly suggests that people vary considerably in their propensity to copy others or find stuff out for themselves. Personality traits such as creativity and curiosity are clearly linked to the ability and willingness to carry out successful individual learning, and these traits vary widely.

There also seems to be a gender difference. Kimmo Eriksson of Malardalen University in Sweden, one of the tournament鈥檚 designers, and Pontus Strimling of Stockholm University discovered this when they carried out a game called explore and collect, in which paired players tried to get the highest possible score among a number of undisclosed options by either uncovering the relative ranks of options for themselves or choosing options already favoured by the other player (). 鈥淲e found that women tend to invest more in individual learning than men, in the sense that they spend more effort on trying out a greater number of unknown options,鈥� says Eriksson.

As well as highlighting the variability in our individual approaches to social learning, the tournament has also shed light on an apparent paradox. Laland and others have found that social learning is widespread in nature, even being used by invertebrates. So what鈥檚 so special about copying in humans?

鈥淪ocial learning is widespread in nature. So what鈥檚 so special about copying in humans?鈥�

Firstly, says Laland, the competition reveals that social learning does not require much brainpower. 鈥淵ou don鈥檛 need any clever copying rules. You can just copy anyone at random,鈥� he says. 鈥淥ther individuals are doing the filtering for you. They will have tried out a number of behaviours and they will tend to perform the ones which are reaping the highest rewards.鈥� That explains why even insects can benefit from social learning. 鈥淏ut,鈥� he adds, 鈥渢o become the winner of the tournament you really have to do something a bit more sophisticated than that.鈥� You have to weigh up the relative costs and benefits of sticking with the behaviour that you have, versus inventing a new behaviour, versus copying others. That requires assessing how quickly the environment is changing, as this gives you an idea of how quickly information will become outdated. Discount Machine was very good at doing just that 鈥� in variable environments it placed a higher value on more recently acquired information and discounted older information more readily.

It is in this ability that humans seem to have the edge over other animals. That鈥檚 not to say we are alone in making these sorts of calculations, though. For example, Laland and his colleagues have found that sticklebacks can do it. First they taught individual fish to expect more food at site A than site B. Then they switched the food around, but the only clue to the deception was that there were now more fish feeding at B than A. It turns out that the longer it has been since the fish checked the sites out for itself, the more it will rely on social information to tell it which site has the most food ().

While this is impressive, humans have a unique talent that allows us to take account of passing time and changing circumstances far more effectively: language. 鈥淵ou can simply talk about what might happen,鈥� says Rendell. Or you can use language to imagine yourself in a different place or time. Rendell suspects this may be what has enabled us to take full advantage of social learning, leading to the huge gap between human culture and the behaviour of other animals.

The tournament has undoubtedly provided several insights into social learning. According to Rob Boyd of the University of California, Los Angeles, a pioneer of social learning research and another of the tournament鈥檚 designers, its big advantage over previous approaches is the level of realism. It entails 鈥渕uch more environmental complexity and more cognitive complexity in the organisms鈥�, he says. Nevertheless, there is room for improvement. Rendell points out that the simulations cannot track particular individuals through time, and that it doesn鈥檛 include formal teaching, a vital part of learning in the real world. 鈥淲e want to explore additional complexities with some more tournaments in the future,鈥� he says.

Before they do that, however, the team has another intriguing idea to pursue. 鈥淲e want to go out and try to explore this in the real world,鈥� says Rendell. 鈥淲e plan to set up an experimental version of this tournament where we get people to play it themselves and see what they actually do.鈥�