> That being said, this _is_ the worst-case scenario for ECJ: symbolic
> regression combines a complex breeding procedure with a fast  evaluation. 
> We tend to use ECJ on much longer evaluations than  this.  But I'm always 
> interested in how we might be able to speed  ECJ's breeding up in the 
> code.  Ideas?
Such hypotetical idea used some years ago in other context, swapping between 
two populations comes mind. Quite extreme is do changes for other and read 
other.

Also was there earlier something about factory pattern and trying own 
resycling impl, this can be outdated opinion but if I remember right own 
re-use pattern can be made quite fast.

t, Harri