might go much faster if it recorded a set of states where ancestors died shortly after and then test all the new candidates against those states. The new candi

might go much faster if it recorded a set of states where ancestors died shortly after and then test all the new candidates against those states. The new candidate might actually survive because its prior history kept it from ever getting into that particular death state, but I think biasing towards designs that don't immediately die in those hard cases is good anyways, since given a long enough run it would likely encounter a similar state. One could co-evolve the test case collection by simulating only the best candidates according to the test cases, and then retaining test cases based on a running score for how well they predicted the actual performance.