Next: Artificial Evolution Re-Discovers Building
Up: Smooth Mutations and the
Previous: Smooth Mutations and the
To compare both types of mutation, we ran 10 iterations of the crane experiment
(section 2.6.3) with four different mutation/crossover strategies.
- 1.
- Original (brittle) mutation + crossover
- 2.
- New (smooth) mutation + crossover
- 3.
- Original (brittle) mutation, no crossover
- 4.
- New (smooth) mutation, no crossover
The result was unexpected (fig. 2.19). We obtained slightly better
performance with the improved mutation procedure. But the loss of performance
when crossover is turned off was extremely large. As illustrated by the figure,
after 150,000 iterations, the fitness reached, averaged over 10 runs, is 16.8
or 20.5 (depending on the mutation strategy) whereas without crossover the figures
fall to 12.9 and 10, respectively.
Figure 2.19:
Comparison
of two mutation strategies (``raw'' and ``smooth''), with or without
use of crossover: Not using crossover results in a dramatic performance loss.
Plots show maximum fitness reached (averaged over 10 runs) throughout the experiment
(left) and at the end (right). Error bars show the standard error of the mean.
|
This result suggests that recombination plays a key role, that of reusing useful
subparts. Recombination, we posit, is responsible for the self similarity observed
between different parts of evolved structures, and thus for the appearance of
higher levels of organization than just the individual bricks.
Next: Artificial Evolution Re-Discovers Building
Up: Smooth Mutations and the
Previous: Smooth Mutations and the
Pablo Funes
2001-05-08