In the Ev model a functional binding site is indicated by the site being rated as having an evaluation greater than the threshold.
The Ev paper does not tell how the threshold for 'mistakes' is determined, which is something I would like to know.
I was puzzled by this but I searched through the paper and found that the terms 'threshold' and 'tolerance' are terms used to mean the same thing in the paper!!! My error, sorry, nobody caught that one before.
For better or worse, things published in the literature are rock solid and cannot be altered. This one is not worth writing an erratum about, though maybe if I republish the paper in a book or something then I could correct it. (I've made a record, but will I remember to look at the record 10 years from now?? ;0)
To answer your question, the threshold is a number embedded in the genome. So it evolves by itself.
I've been curious recently as to whether or not the evolution would take place with the threshold set to zero. So I made a cheap version of the program that does this and found that it still evolves.
For Ev parameters that give Rfrequency = 3.0, the final Rsequence for the normal ev at generation 1000 was
1000 r 3.66567 +/- 0.55357
whereas for the program with the threshold forced to zero I got:
1000 r 2.90320 +/- 0.55357
I worried a little that this was low so went another 1000 generations:
2000 r 3.45200 +/- 0.55357
so within this crude estimate, there is no difference (note also the error range). It's just in there for historical reasons - we were thinking about perceptrons (a simple neural net) at that time and they have a threshold, so it was natural for me to do the same for Ev. Apparently it isn't very important.
origin: 2005 Sep 20
updated: version = 1.00 of zerothreshold.html 2005 Sep 20
Viewing Files Accessibility