[Martin Taylor 2009.01.11.23.07]
[From Rick Marken (2009.01.08.1950)]
Actually, as soon as you say that, you put a conditional on the Bayes
analysis "Given that the data consist of the correlations". I'll work with
that, but I kind of hoped there might be some access to the four tracks,
There are only three tracks and they are certainly available if you need them.
Four, no? One for each of the three characters, and one for the mouse
(three disturbances and one reference).
I found that the reason I was getting no effect of the mouse was that I
was holding down the mouse button (which seems to me most natural). On
the Mac Firefox, I get what you say – jumpy. On Windows Firefox,
running under Parallels 4, it’s smooth. I didn’t manage to use Ubuntu
Firefox yet, because it seems I still don’t have Java properly
installed. However, having one working version should be enough.
So where would I find the track data? I assume it’s in a file
somewhere, not accessible from within the program.
Martin
[From Rick Marken (2009.01.12.2115)]
Martin Taylor (2009.01.11.23.07)_-
Rick Marken (2009.01.08.1950)--
There are only three tracks and they are certainly available if you need
them.
Four, no? One for each of the three characters, and one for the mouse (three
disturbances and one reference).
What I call the tracks are the time varying positions of the three
characters. There are actually four relevant pieces of data: the three
disturbances to each character (d1, d2 and d3) and the mouse position
(m). You can then calculate the three tracks of the characters as
d1+m, d2+m and d3+m.
So where would I find the track data? I assume it's in a file somewhere, not
accessible from within the program.
The data (d1, d2,d3, m) is available only in arrays in the program.
But I don't see why you need any actual numbers to show me how to use
Bayesian statistics to improve detection of the controlled character.
Just assume that the three disturbances are sine waves (each slightly
out of phase with the other and m is approximately the negative of the
disturbance to the character under control. I am calculating three
correlations: d1 with d1+m, d2 with d2+m and d3 with d3+m. I take the
lowest correlation (by some threshold amount) are the one for the
controlled character. So how can Bayes help?
Best
Rick
···
--
Richard S. Marken PhD
rsmarken@gmail.com
[Martin Taylor 2009.01.13.09.33]
[From Rick Marken (2009.01.12.2115)]
Martin Taylor (2009.01.11.23.07)_-
So where would I find the track data? I assume it's in a file somewhere, not
accessible from within the program.
The data (d1, d2,d3, m) is available only in arrays in the program.
But I don't see why you need any actual numbers to show me how to use
Bayesian statistics to improve detection of the controlled character.
I could do it in the abstract, but it would be nicer to have an actual
example or three as a test case. Without having the tracks, the only
“real” data available are the correlations, and one of the points of
the exercise is to determine whether the Bayes approach might suggest a
better approach. However, if I must use only the correlations, I see
them flickering around on the display, but how would I find the record
of their time course?
Just assume that the three disturbances are sine waves (each slightly
out of phase with the other and m is approximately the negative of the
disturbance to the character under control. I am calculating three
correlations: d1 with d1+m, d2 with d2+m and d3 with d3+m. I take the
lowest correlation (by some threshold amount) are the one for the
controlled character. So how can Bayes help?
It is conceivable but I think rather unlikely that a Bayesian approach
might help if the data available are restricted in this way. When I
suggested that it might improve the ability to determine which
character was being controlled, I was presupposing that the Bayesian
analyst would have the opportunity to use the same data as you use in
doing the correlational analysis. Even then, I said that the answer
might be to discover that your algorithm was in fact optimal (which
ought to be reassuring, if it so happens). But if two approaches don’t
have access to the same data, no analysis can tell which might be
better, or how they compare in relative efficiency.
I’ll try to do it in the abstract, then.
Martin
[From Bill Powers (2009.01.13.0752)]
Martin Taylor 2009.01.13.09.33 --
When I suggested that it might improve the ability to determine which character was being controlled, I was presupposing that the Bayesian analyst would have the opportunity to use the same data as you use in doing the correlational analysis. Even then, I said that the answer might be to discover that your algorithm was in fact optimal (which ought to be reassuring, if it so happens). But if two approaches don't have access to the same data, no analysis can tell which might be better, or how they compare in relative efficiency.
In line with the post I just sent, I suggest that there is a systematic way of determining what character John is controlling. We ask, "John, which one are you controlling?"
As long as we require the condition that we must do this statistically, we will be working with one hand tied behind our backs. If you doubt that my method is better, you could of course determine, like a good frequentist, how often the asking method is correct in comparison with the ultimate determination by the statistical method, but once you had satisfied yourself that my method was almost always faster in reaching the right conclusion, you could drop the statistical method.
I advocate using the statistical method when there is no better method available, but only then.
Best,
Bill P.
[Martin Taylor 2009.01.13.10.55]
[From Bill Powers (2009.01.13.0752)]
Martin Taylor 2009.01.13.09.33 --
When I suggested that it might improve the ability to determine which character was being controlled, I was presupposing that the Bayesian analyst would have the opportunity to use the same data as you use in doing the correlational analysis. Even then, I said that the answer might be to discover that your algorithm was in fact optimal (which ought to be reassuring, if it so happens). But if two approaches don't have access to the same data, no analysis can tell which might be better, or how they compare in relative efficiency.
In line with the post I just sent, I suggest that there is a systematic way of determining what character John is controlling. We ask, "John, which one are you controlling?"
As long as we require the condition that we must do this statistically, we will be working with one hand tied behind our backs.
What do you imagine to be the point of Rick's "Mindreading" demo? Why doesn't just ask the person to click a button for which character they intend to control. It would be a much easier mindreading job, wouldn't it? I'm sorry you think Rick is too stupid to see that he is working with one hand behind his back.
Martin