BBC Atricle: 'Two simple rules' explain sheepdog behaviour

[Maxi]

There is a vast scientific literature out there. Unfortunately, not all of it is high quality and some is frankly quite dodgy. There is a big difference between high quality studies that use good/robust observation/experimental design and those that do not.

 

Not surprisingly, it is often very difficult for someone without the relevant scientific background to be able to separate the wheat from the chaff. However, what some scientists are also unwilling to acknowledge is that their knowledge of the subject they are investigating may also limited (as I personally believe has happened to the researchers in the original sheepdog herding study that started this discussion)

 

As Mr McCaig so rightly says says "I find the interaction between sheep, sheepdog and shepherd astonishingly complex. Some humans understand this interaction brilliantly though few are able to articulate their knowledge."

 

Unfortunately, IMO the apparent lack of knowledge of the researchers about sheep-dog-handler dynamics has meant that they have over-simplified the process to gather and drive, when anyone who works with stock knows that this has taken the reductionalist approach too far.

 

In addition, as well as quality of the data/experiment, it is also necessary to distinguish between true observation (fact) and interpretation/extrapolation of that fact.

 

Again to give an example from the sheepdog perspective, I personally find that pressure - release techniques work well when I train my dogs. That is my observation. I have adapted/refined my approach (i.e. done experiments) to improve this.

 

However, some other individuals might then like to extend these observations to suggest that pressure - release techniques only work because they use the hierarchy/dominance/pack theory and work because a dog has some iinstinctive feel because of his evolutionary past. Comments like this cannot be proved. It is is only what that individual interprets and extrapolates. (Similarly, it is probably not possible to prove any other interpretation).

 

So yes, Mark (post #6) was right, I should not just have commented on the popular press report without reading the original paper. However, when I did as he recommended, I realised that the poor quality experimental design actually meant that my original statement was probably more correct than I had even intended. Furthermore, the questions that Mark himself posed in post #6 indicate how ridiculous their claims about the future potential use actually is.

 

So yes, I am fully supportive of good robust science.. However I am also well aware & very wary that a lot of incorrect/misleading 'science' enters the public domain . It is then then believed to be an absolute truth because it came from a 'scientific report'.

[Donald McCaig]

 

Dear Doggers,

My friend and working scientist Mark disagreed with(Paraphrase) "Dogs can detect bowel cancer ay 80% accuracy when other methods are too late. Scientists are interested but wish to replace the dogs with machines.

 

 

"80% accuracy does not really meet current clinical diagnostic methods (better than nothing for now).

a. I'm sorry for making you go through surgery (or unnecessary invasive testing) on your bowl; there was no cancer. You are one of the 20% of the patients our dog does not diagnose correctly.

b. I'm sorry your cancer is now late stage; I guess our dog missed your cancer during early screening.

 

If dogs can smell odors produced by cancer than there are chemical markers unique to cancer for which a more accuract test can be developed. Dogs can be used to help guide researchers to those chemicals and a more accurate test."

 

First, no diagnosis or medicine guarantees, 100% success. 80% when that's better than anything else is better than anything else.

 

And we don't know how dogs detect cancer. Perhaps some slight behavioral variation gives it away. What we know is that (some) untrained dogs can. As a dog person faced with this problem, I'd wonder not: "Can we make a machine?" but "Can we find better prospects and train it for better results?"

 

Donald McCaih

[alligande]

Thank you for an interesting conversation ladies and gentleman as a non-scientist I have found your interpretations a different way to look at what I thought was just a bit of BS research and given me something to ponder next time I see/read a short segment on a weird bit of science.

[Pippin's person]

Maxi (and others), I completely agree that this article doesn't capture the nuances of the various dynamics involved in sheepherding and that it's quite possible, likely even, that the researchers don't fully understand all that is involved. In addition to their real world data involving only one dog, one human, and one flock of sheep, the size of the field was pretty small (150 meters) and the tasks involved pretty basic.

 

Those facts, though, don't necessarily make the conclusions drawn in the article inaccurate. The article set out to mathematically model (in contrast to previous models which were strictly derived from theoretical perspectives on animal behavior) why a dog can move a large mob of sheep. They don't claim to capture the whole array of things going on when working sheep with a dog (despite how the popular press presents it).

 

Yes, they built the model first and then looked to see how well the model fit a set of real-world data--that's a perfectly reasonable (e.g. acceptable, standard, etc.) way to approach that component of the question they were asking (indeed, partly their question was "does our model fit real world data"?). They found that it fit pretty well. (ETA: For the very limited task it was modeling)

 

What they were modeling was how one agent keeps a group moving--they found that the ability for one agent ("the shepherd", by which they mean either the computer simulated shepherd or the Kelpie) to successfully affect (e.g. move to a pre-determined spot) the other agents (the simulated or real sheep) was tied to various properties of angle and distance of the "interacting agents".

 

They also show that the side-to-side motion used by the "shepherd" (e.g. computer simulated shepherd or Kelpie) is a direct result of shifting between "collecting" and "driving" rather than some kind of hard-wired behavior in which the side-to-side motion comes first (that's how this model differs, they claim, from some other models that have posited the side-to-side motion as what is moving the sheep rather than as a consequence of shifting between "collecting" and "driving").

 

I hadn't read that part of the study as carefully yesterday, but to me, that's pretty interesting--that you can model wearing not as the primary behavior but as a result of shifting between keeping the sheep together and moving them forward. Maybe that seems obvious but previous models have apparently assumed the side-to-side motion as primary rather than derived.

 

They don't claim that by having demonstrated that their algorithm matches what the Kelpie did that they have captured the essence of sheep herding with a dog. (again, contra to how this study was presented in the popular press)

 

Personally, I think there are really interesting questions that can arise from this study--including exactly how well the model fits other types of dogs or whether you could model other aspects of herding and if so, what that would mean. I suspect that in a similarly controlled circumstance (a trained dog asked to move a group of sheep to one corner of a 150 meter field), you'd probably get very similar results--but of course, that's an empirical question.

 

ETA: The generation of new questions, as a result of small steps (sometimes wrong ones), is one important mechanism for how our knowledge base expands.

[Mark Billadeau]

Since Donald and I are drifting a bit off topic I'll just add this link for those who are interested:

Canine scent detection of human cancers: A review of methods and accuracy (full text)

 

Donald, note the researchers state scent as the means of detection; and while astounding and fascinating to watch in action dogs are not 100% reproducible in their execution of the trained tasks. Not a bad thing for most scent tasks, perhaps not totally desirable for cancer diagnoses.

[Debbie Meier]

I don't believe that a robot can read intent nor process other environmental conditions that would make the livestock want to go or not go to a particular place.

 

Each time I watch the video of the remote controlled car moving cattle I laugh, if not for the person getting the car out of trouble it would get tromped by a curious cow that decides that it's enough of a threat to kill but not so much to fear.

 

How do you program a touch of self preservation into a robot?

[Maxi]

Although I will continue to stick to my guns about this particular research paper, I fully respect that others have a differing view and opinion on it.

 

I just would like to say that it is a real pleasure to be part of a discussion forum where indiviuduals are allowed to express theirs divergent opinions.

[Mark Billadeau]

So who might want to use programs for automnomous robots which move like a Border Collie to collect and drive away animals?

 

Clear Flight Solutions

[geonni banner]

Well, I don't know about collecting, but airports could use them to keep runways clear of gulls, geese, rabbits (a problem in some places.) The noise on a runway can be overwhelming. Robots wouldn't care. But dogs have very sensitive ears.

 

What about game parks or reserves? To keep the game in, or drive them back if they stray onto farmland or into villages where they destroy crops and injure people. Even elephants can be easily frightened by novel sounds or sights. And an elephant away from a protected area is in even greater danger from ivory poachers that one on a patrolled reserve. A robot could be armed with flashing lights, strange sounds or smoke bombs. Even large predators would retreat from such things - and even if they didn't, better a chewed robot than a chewed villager or game park employee. They could be linked with patrolling drones to alert them of boundary issues.

 

They could be used in paces like national parks to discourage bears from hanging out in campsite areas or garbage dumps.

 

Orchards and some field crops suffer depredation by deer at certain times of the year. A robot could be detailed to scare/ drive them away at night. Deer would likely be leery of something that moved like a predator.

 

Waterproof ones could drive sea lions from beaches that are heavily used by surfers, swimmers, and other "recreators."

 

In some places, protected nature reserve areas are intruded upon by herdsmen driving their livestock to feed when their own pastures are skimpy. The Gir lions in India are suffering dwindling numbers because of this. The herdsmen's cattle destroy forage and cover that the lions and their prey species need to eat, hide and hunt. Certainly a dog would be better for this, but many drovers would have no compunction about killing a dog. A robot might be harder to "kill." And it comes with training (read programming). Once production models were available they could be manufactured relatively cheaply. And if paired with surveillance drones, there could be a sound/video record of interference with them.

[geonni banner]

How do you program a touch of self preservation into a robot?

It can be done. One of my favorite stories about the making of the "Lord of the Rings" trilogy of films was about the hundreds of CGI characters generated to fill in large battle scenes. They were programmed to fight, but there were problems with them, so they gave them simple programs that gave them an option to fight or run when overmatched. When the program was finished they did a simulation run and all the "fighters" ran away! So they had to do some tweaking...