RSS 2.0 Feed

» Welcome Guest Log In :: Register

Pages: (14) < [1] 2 3 4 5 6 ... >   
  Topic: Evolutionary Computation, Stuff that drives AEs nuts< Next Oldest | Next Newest >  
Zachriel



Posts: 2723
Joined: Sep. 2006

(Permalink) Posted: June 25 2009,08:41   

The calculation of "working fitness" is seemingly broke. From Mendel's Accountant:

      do i=1,total_offspring
          work_fitness(i) = work_fitness(i)/(randomnum(1) + 1.d-15)
       end do


We can test this by taking a series of fitnesses k from 1.001 to 2,

 For k = 1 To 1000
      Cells(k, "a") = 1+ k / 1000
      Cells(k, "b") = Cells(k, "a") / Rnd
  Next k


This is a typical result:

9 Average
31 St.Dev.
362% Relative St.Dev.
1.04 Min
533 Max

The original distribution of k has a Relative St.Dev. of 19%. It's worse for fitnesses distributed between 0.5 and 1.5 or 0.5 and 1. (Just like the phylogenetic fitness, the calculation is not normalized. And why ÷Rnd^1? Why not ÷Rnd^½ or ÷Rnd^¾?) This single operation eliminates the vast majority of the signal from genetic or phylogenetic fitness.

A more reasonable calculation is Roulette Wheel selection.

I have a working version of Gregor's Bookkeeper. I'll post on that in the next few rotations.


-
Fixed a problem.

--------------

You never step on the same tard twice—for it's not the same tard and you're not the same person.

   
  419 replies since Mar. 17 2009,11:00 < Next Oldest | Next Newest >  

Pages: (14) < [1] 2 3 4 5 6 ... >   


Track this topic Email this topic Print this topic

[ Read the Board Rules ] | [Useful Links] | [Evolving Designs]