Learning in different search landscapes

I just published, together with Claudio Tennie and Alex Mesoudi, a new paper in Royal Society Open Science: Social learning solves the problem of narrow-peaked search landscapes: experimental evidence in humans.

The paper describes an experiment that is based on some modelling work that we did few years ago. The basic logic of the model was that learning could “be represented as a search for an optimal behaviour […] constrained by different kinds of information” (citing from here), and the only difference between social learning and individual learning was that, in the former, these information were provided by the behaviour of conspecifics. This seemed to me, already then, a quite advanced way to see the distinction social/individual learning, at least in respect to “standard” cultural evolutionary models, but I never focused on that too much later on. Anyway, the models included different “search spaces”, and the main result was that hi-fi social learning (then “imitation”) was useful in narrow (or “peaked”) search spaces. These search spaces are characterised by the fact that behaviours close to the optimal one not only are not good, but, especially, for the functioning of the model (and, hopefully, of real learning) they do not provide feedback on how close one is to the optimal behaviour (see e.g. the red line in the figure below).

f1-large

In subsequent papers I usually refer to this as the “Windsor knot” problem. Tying a Windsor knot is a difficult task that requires hi-fi social learning, and at least part of the difficulty derives, according to this logic, from the fact that performing correctly the majority of the actions does not generally create an almost-good Windsor knot. In other words, there is not a way to search individually in the Windsor knot-search space.  In smooth spaces (see e.g. the blue line in the figure above), instead, individual exploration is easier, and tasks that can be described by these search spaces should be solvable through lo-fi social learning, or individual learning.

14010915570_6cea8e8367_z-e1423124404414

Fast-forwarding a few years, I read Alex Mesoudi’s virtual arrowhead papers (see e.g. here or here). One of the findings of these papers was that, while social learning was more effective than individual learning, it was underused by participants. My first reaction was: I know why! In fact, the search space was multimodal (i.e. more than one peak) but smooth. I then proposed to run the same experiment (in Birmingham, thanks to Claudio), but “narrowing” the search space, testing the original model’s hypothesis (individual learning should be more difficult in narrow-peaked search spaces, but social learning should solve the problem) and also whether narrow-peaked spaces would have increased the number of participants using social learning.  Long story short, the first hypothesis was correct, but not the second (not surprising a lot more about this in the paper). There was a difference between the narrow and the smooth condition regarding the use of social learning, in the right direction, but not statistically significant – not sure if I can say this, but I still stubbornly think I am right, and that if there were more participants one could see this effect…

In any case, the experiment is important because shows how different search spaces can radically change the effect of social and individual learning. An explicit attention to the search space structure of the experimental task used might greatly improve our knowledge of when and how social learning is necessary and when not. Also, I believe that this could be a way to look to past experiments, and see whether the search space they used could have had an effect on their results.

Royal Society Open Science has been an excellent outlet for the paper. Platinum Open access (i.e. free for both readers and authors, but this is just for the first period), and with swift, and especially, open, reviews. The reviewers could reveal their identity (one did), and all is accessible to everybody here. Surely, a step in the right direction.

References

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s