I did a git clone last night and carefully read through the starter kits for Common Lisp, Java, and Python. A long time ago, in my book "C++ Power Paradigms" the last example was using a genetic algorithm to train a recurrent neural net for the Sante Fe Ant Trail problem. My original idea was starting with weights represented with relatively few bits, and adding less significant bits slowly during training. My hope was that I would quickly find reasonable areas of weight values so the search for good values with a reasonable number of bits per weight would not take too long. After I wrote that John Koza told me that I had an interesting idea but probably not useful. That said, I am tempted to brush off my old idea and try it again.