Thanks for your message. I modified the seeds for the Mersenne Twisters and
now, both trees evolve. Thus, you can relax since it shouldn't be a bug.
However, the performance isn't that satisfying yet. Thus, I tried your
second approach and it delivers quite nice results. The only downside is the
very complex tree that is created. I will try to use some of the parsimony
functions (like the lexikographic selection). Otherwise, I'm gonna update to
the fitness function which should penalize too large trees. In case that
doesn't work either, I'm gonna have a look at the CGP framework which seems
to be able to limit the program bloat effectively.