Its fine that we can take 1000 instance generations and pick the most successful after a trial, but that has problems, what if you only have one instance? (say youd like it to learn off a user, not a program)

My idea, is maybe i can get an ANN to evolve in one instance.

Have a running success rate, that a mediator program modifies.

Add the success rate (a very small number), success can either be positive (reinforcing) or negative (inhibiting), from active synapses, only cells that fire get their output synapse permanences updated. this rates the synapses for how generally successful they were.

then randomize the least successful synapse each frame, instead of the whole network.

So each frame, one synapse gets randomized.

Its a huge stab in the dark, what do you think?

?

A stab in the dark? Not really. Sounds like backpropagation.

There is one really good backprop tutorial. I needed 3 days to figure out what is happening, then I found it. Just from one read I implemented ANN library.

http://clemens.bytehammer.com/papers/BackProp/index.html

Good luck :)