[From Bill Powers (941209.1050 MST)]
Bruce Abbott (941208.1440 EST) --
Yeah, I tried out the equivalent by pressing the space bar to tumble e.
coli, based only on watching the dNut display. Not pretty, but it
works (my e. coli was all over the map, but did tend to "find" the
nutrient source).
The reason it seems to work poorly is that the programs we've been using
work too well. When a tumble is initiated by ANY negative value of dNut,
however small, there are many tumbles that occur while the spot is
within a few pixels of the same position, so all we see is the effect of
the final tumble that got dNut positive again, and this happens as fast
as the program can iterate. Doing it by hand (particularly when watching
numbers change on the screen), a longer time is needed to see that the
change is in the wrong direction, so the spot can move appreciably in
the wrong direction before the human operator can initiate another
tumble. Furthermore, because of this delay, we see the effect of _every_
tumble, and thus all the path segments that are going the wrong way. The
human behavior is more like that of the real E. coli than is the final
behavior of Ecoli4.
When an analog representation of dNut (or more simply, the radial
distance to the target) is used, human performance is better. But I find
that if I try to make corrections less than about 1/3 to 1/2 second
apart, I can't stop the pressing soon enough to prevent initiating a
tumble when the variable is changing the right way. So the spot ends up
moving for a minimum of a half-second or so between tumbles. If you put
a similar constraint on initiating tumbles in the model, you'd see the
same behavior.
ยทยทยท
-----------------------------------------------------------------------
Best,
Bill P.