Practical Neural Network Recipies in C++

144 Chapter 8

if (ftemp>

1.e-20){

tmult = avgfitness / ftemp ;

tconst = -minfitness * avgfitness / ftemp ;

}

else {

tmult = 1.0 ;

tconst = 0.0 ;

}

}

The scaling factors have been computed. Do the scaling now. The

truncation at zero is theoretically unnecessary, as we avoided negatives

when we computed the scaling factor above. However, floating point

problems can sometimes still produce a 'negative zero'. In deference to

possible user modifications which DEMAND nonnegative fitnesses, it is

good programming practice to enforce this.

avgfitness = 0.0 ;

for (individualo ; individual<popsize ; individual·^) {

fit = tmult * fitnesspndividual] + ...