SoylentNews is people
SoylentNews
The fact that animals can use tools, have self-control and certain expectations of life can be explained with the help of a new learning model for animal behaviour. Researchers at Stockholm University and Brooklyn College have combined knowledge from the fields of artificial intelligence, ethology and the psychology of learning to solve several problems concerning the behaviour and intelligence of animals.
Animals are often very effective; an oystercatcher opens mussels quickly, a baboon takes every opportunity to steal food from tourists or a rat navigates with ease between the bins in a park. Previously these behaviours have been considered to be inherited instincts, even though it is well known that animals have great learning abilities. Researchers from Stockholm University and Brooklyn College have now created an associative learning model that explains how effective behaviours can arise. This means that an animal does not only learn that the last step of a behaviour chain, the one that is rewarding, is valuable. An animal can learn that all steps towards the reward are valuable.
"Our learning model may also explain how advanced behaviours are created at an individual level. Behaviours like self-control, chimpanzee tool use as well as other phenomena like animals having certain expectations of live", says Magnus Enquist, professor of ethology at Stockholm University. "Similar models are used in the field of artificial intelligence, but they have been ignored in animal studies."
Since the 1970s it has been known that animals weigh the cost of a certain behaviour against the profit and that they, to a high degree, make optimal decisions, which is assumed to be genetically determined. The research group's new model deals not only with learning, it also takes into account the idea that what animals are able to learn can be genetically regulated.
Stop. Teaching. The. Animals. Skills.
(Score: 0) by Anonymous Coward on Friday December 02 2016, @02:31AM
From the paper:
I haven't read the entire paper yet, but I have to question the expertise of anyone studying this who leaves out Thurstone:
http://search.proquest.com/openview/a75e05d892575ecdf471807d137c4c64/1?pq-origsite=gscholar [proquest.com]
He already explained the sigmoid curve very simply, long ago.
(Score: 0) by Anonymous Coward on Friday December 02 2016, @03:08AM
All research must be done within the proper social network and certainly incompetent is anyone who neglects to name-drop the name of your hero.
(Score: 0) by Anonymous Coward on Friday December 02 2016, @03:45AM
Really they don't compare the performance of their model with any other reinforcment learning models. Thurstone's is just one of the earliest, the simplest to conceive, and they don't even mention him.
(Score: 3, Interesting) by khallow on Friday December 02 2016, @03:45AM
I find it more interesting that they use at various points explicit thermodynamic models (see equation 2.2, the variable beta "regulates exploration". It's actually the inverse of an temperature equivalent variable in a standard ensemble model) for various parts of the learning model.
(Score: 0) by Anonymous Coward on Friday December 02 2016, @04:28AM
No, there are other models that deduce the sigmoid curve from basic postulates. It isnt something arbitrary.
(Score: 1) by khallow on Friday December 02 2016, @06:51AM
No, there are other models that deduce the sigmoid curve from basic postulates. It isnt something arbitrary.
Nor is the step function arbitrary either.
(Score: 0) by Anonymous Coward on Friday December 02 2016, @01:20PM
Yes it is. In the paper they use a softmax function to decide which behavior is chosen, because it is "simple". Everything else follows from that. The model is arbitrary at its foundations.
(Score: 2, Interesting) by khallow on Friday December 02 2016, @08:26PM
I'll note too that the step function is the far more common model of learning in the real world with most human education based on this. You don't get 80% of a high school diploma when you leave high school. You either get a diploma or you don't.
Similarly it is widely prevalent in culture too, when some movie shows its heroes learning some important task or personal truth, they often montage it with the end result being the heroes having learned the task or truth. In computer games, you learn a combo/trick or not with quantifiers used to describe how much your learning deviates from the step function ("I know how to barrel roll, but I'm not very good at it.").
In any case, due to the simpler math and the widespread application of the step function to learning, it's neither surprising nor unjustified to use the step function as they do in this research.
(Score: 0) by Anonymous Coward on Friday December 02 2016, @10:29PM
If the point was just to describe learning curves, or simulate them, that would be one thing. They are claiming much more:
They are claiming their mode may explain learning curves. And really I am coming across much more negative than I am. I actually really like this paper, I would put it in the top 0.01 %. Still, I don't see why they don't include comparison to other models (such as Thurstone's). Thurstone's in particular, while definitely simplified heavily in terms of assumptions, is not based on any arbitrary functions that appear out of nowhere. Rationally deduced models like that are much easier to extend.
(Score: 0) by Anonymous Coward on Friday December 02 2016, @11:58PM
pdfpls?
1930 so not like copyright should matter.
(Score: 0) by Anonymous Coward on Saturday December 03 2016, @02:07AM
It should be free at that link (you need to enable js to see it though...) It isn't for you?