A.3 Scaling Method

If the normal selection ruler is used during the GA computation (pselecti = fi/img/sum.jpg (262 bytes)f ) in later generations, when the population average fitness gets close to the population best fitness, the above-average members are overwhelmed by average members and under-average ones. The computation becomes more a random walk. In this case fitness scaling can help. The linear scaling looks like this :

f' = af + b.


As a solution the following alternative was chosen.

All genes under favg are degraded by the Deg value. The Deg is a multiplicator between <0,1> specified by user.

The possible implementation:

void CalculateScaling (TPopulation &pop, double deg, double &avg, 
                       double &a, double &b)
   double total=0;

   for (uint i=0; i<pop.GetSize(); i++)
   if (pop.GetSize()) avg = total / pop.getSize();
     else avg = 0;
   if (avg<1) a = (deg*avg - 1) / (avg - 1);
     else a=1;
   b = 1 - a;

TFit ScaleFit (TFit from, double deg, double avg, double a, double b)
   if (from <= avg) return (deg*from);
   return (a*from)+b;

Programmer's Guide Project Antares