🚀 go-pugleaf

In article <2360@blue.cis.pitt.edu> genetic+@pitt.edu (Dr. Dave) writes:
>
>I say "should" above because, as far as I know, nobody has done a systematic
>study on this.  However, it seems to be not uncommon, at least when looking at
>combinatorial problems, to see cases where SA is faster, and running a lot of
>different SA runs will find you solutions just as good as the best GA runs, 
>but the *average* value of the GA solutions is better than the average value
>of the SA solutions, and the quality of the GA is less sensitive to parameter
>settings (pop size, mutation rate, encoding and operators, etc.) than the SA
>is to its parameters (cooling schedule, settle time at a given temp., etc.).
>
>I'd be happy to be corrected about any of the above, if anyone out there has
>different experience/intuition...
>
>Dave Tate

The GA is insensitive to encoding due to the way of crossover operation.
However, we also can say it IS sensitive to encoding once the encoding
method is chosen.  Because any encoding of numbers will have some
bits is significant than other bits, because such inhomogeneous of
bit encoding, the crossover will produce offsprings in the 'neighborhood'
of the parents.  The radius of the neighborhood is random and variable
depending on where the crossover points are cut.  I haven't figured
out the statistic nature of such variable radius w.r.t. particular
encoding.  The recent 'uniform crossover' method is more of multi-bit
mutation than crossover.  

Another thought is that GA is not quite working if the error surface
contains long nerrow stiff valleys.  Usually, GA will stuck, not
matter how you tune the parameters(those rates and pop size).  Once,
it stucks(converges prematurely), it can never get out by itself because
of the similarity among pops.

SA also has the problem of stucking in some local extremes.  However,
the process is dynamic due to the temperature, if let it run long enough,
it 'might' be able to jump out from the stagnation point.  The difficult
of SA is to determine the schedule and corresponding initial temperature
for a particular problem.  The error surface with long nerrow stiff valleys
, SA can perform  better than GA.  This is equivallent to say that 
mutation with certain distribution is preferable than crossover with
random cut to such error surfaces.  However, if the area of extremes
is not narrow but stiff, GA may work better than SA.

Any comments?  I would appreciate if any.

Boris