I''m currently trying to write a simple evolution of NNs for a small 2D battleships, flying around. But it doesn''t seem to work because probably I don''t understand how to optimally specify fitness result. I let them drive and shoot, and at the end of a round, I use such a formula: fitness = damage done*(health left/health max); But in, say generation 4500, some didn''t even fire their guns, being stuck in the corners. I''ve got about 20 inputs with vectors of enemies, of closest bullet(to dodge it), distance to the borders, last movement and my out is 6 values: left-right (angle), forward, backward (velocity=forward-backward). Fire(>0.5) and Shield (>0.5). Can anybody recommend something? I''ve sucessfuly evolved a simple - go to point NN(With left-right, forward e.t.c). But then fitness was easy - more objects to gather-better NN. Plus 2inps, 2outs instead of 20-6 ;-(

the probability of your NN will fire is rather low, so try mutating it alot, and do not allow any NN with fitness 0(in other words, any NN which doesn´t do damage) "reproduce"...Just rerandomize them...

Then keep on randomizing until you get a number NN''s that does damage, and then starts breeding new NN''s....

I am looking to do something with NN''s too, but not right now