next up previous
Next: SMART Operator Application Up: Discussion Previous: Discussion

SMART Operator Recombination


One of the most obvious issues of SMART operator evolution is ``Who is recombining the SMART operators?'' As was mentioned in section 3.4, the results of this chapter have used the simplest solution: random recombination. Isn't this exactly what we are trying to avoid by having the SMART operators? Yes, at a meta-level. In general there are three solutions to this meta-recombination problem: random recombination, self-recombination, and Meta-SMART recombination.

Random recombination is a reasonable option. This chapter does not claim that random recombination in PADO (or GP in general) does not work. Probably the SMART operators would learn a little faster if the recombination used on them was more intelligent but there is a price.

Self-recombination seems at first like a good idea but we have found it (empirically) to be lackluster. The problem appears to be that if the SMART operators' fitness is based on their ability to recombine both the main population and each other then they are evolutionarily ``pulled'' in two different fitness directions. If their fitness is based only on the main population recombination results, then they are evolutionarily ``pulled'' away from the sort of intelligence needed to recombine each other. We hypothesize these results occur because the syntactic needs of the two populations, even on the same generation, are very different.

Meta-SMART recombination is the third possibility. Why not make Meta-SMART operators for the SMART operator population? This strategy is reminiscent of [Schmidhuber 1987]. In our opinion this simply shuffles the problem around without solving it; what happens to the Meta-SMART operators?

Eric Teller
Tue Oct 29 17:04:55 EST 1996