The idea is simple. We have all seen image recognition. Recognition of different things like faces, animals and things. I have a feeling that recognition by machine learning is the basis of many things like set generation.
The idea is that you can’t generate a new set only by optimization. That is. If you optimize a bicycle for instance you won’t get to a car.
Optimization is limited. What you need is some new information. Here is where recognition by machine learning comes in.
Say you have a new set of 100 optimized machines and machine learning software that has learned what the machine type looks like from previous series.
The recognition software could then set a value on each of the new 100 machines. Say from 0 to 100%. This value would indicate how well it was recognized as the machine type.
The idea is then to match machines with a low recognition value but with a high performance value.
These two pieces of information could then indicate new information.
A beginning of a new set.
Another idea would be that you keep those machines that show an interesting feature but perhaps have a lower performance value than what you would normally find acceptable. To let it develop to something better later on.