Here is an interesting couple of articles from Steve Jurvetson about weighting up the benefits of designing your algorithms or letting them evolve. Read the comments as well as they are pretty good.
More work is required in using Bayesian reasoning on evolved algorithms, essentially using the power of Bayesian to determine the set of information which you don't know and continuing working on whittling that set down.
Directed evolution would potentially help as well, dependent on understanding to the point of exhausion what your assumptions are in setting the direction or guidelines or rules (framework)
Have Fun
Paul
