Mutation of constant factors should not simply add a normally distributed value but a value sampled from a mixture of Gaussians

[gkronber]

Now mutation of constants and variable weights is:

f = f + x
x ~ N(0,sigma),

where sigma is a parameter supplied by the user.

This should be changed as suggested by maffenze to the following more complex adaption routine that is arguably more robust.

f = f + y
y ~ N(0,s)
s ~ U(0, s_max),

where s_max is a parameter supplied by the used. For variables s_max should automatically set to:

 s_max = s_user * Var(x_i),

Where s_user is a user supplied parameter and Var(x_i) is the variance of variable x_i over the observations in the training set.

[gkronber]

I suggest we use the switch to the new version (allowing behavior changes) to implement this change.

An quick experiment should be made first to check that algorithm performance doesn't deteriorate.