The simplest addition to the SMART operator paradigm is probably the concept of seeding each new population of SMART operators (at generation 0) with a population that has already evolved in the same or a similar domain. Surprisingly, even for the same domain, this has not worked well experimentally, though we have so far been unable to find a satisfactory theoretical explanation. In fact, even the most fit SMART operator from generation 100 is often less good on generation 25 (when seeded) than the most fit SMART operator that evolved there from scratch.
More generally, the question may be asked, ``What happens when the co-evolution of the SMART operators is discarded completely?'' In other words, it might be possible to evolve a set of SMART operators that would generalize across runs, across domains, or even across generations. Suppose that a set of SMART operators were co-evolved on domain X and that the population of SMART operators after generation Y was saved to a file called SMART.X.Y. To generalize across runs would mean that SMART operators (SMART.X.Y) could be used on each run in domain X and that on generation Y the SMART.X.Y population would be used for recombination. To generalize across domains would mean that SMART operators (SMART.X.Y) could be used on each run in any domain (not just domain X) but on generation Y the SMART.X.Y population would be used for recombination. To generalize across generations would mean that SMART operators (SMART.X.V) could be used on each run of domain X but on each generation Y, the SMART operator population SMART.X.V for some fixed V could be used. The term ``could be used'' in this paragraph means not only ``was worth using because it generally outperformed the random operators,'' but also ``was worth using because it generally outperformed the SMART operators that would have co-evolved to that generation.'' The summary of our experiences has been that while canned (i.e. saved and recovered) SMART operators generalize well across runs, they do not generalize well across domains or generations. In practice, we have found no seeding paradigm worth the effort. However, these results are little better than anecdotal; our on-going work in this area may still bear fruit.