On 28 March 2011 22:25, Raul Miller <[email protected]> wrote: > Note that, personally, I do not know what you are trying to accomplish. > However if you are trying to build a list where each element is double > > No, it'll be those random numbers again. I'm trying to simulate the road that my neural network is gonna try to drive along. It should wind back and forth more or less like real roads. The list I'm making is a set of values for the curvature at each moment in time (the car's speed is constant for now.) So each value is a small random change to the previous one, with a bit of tuning if I can be bothered. I'm thinking of choosing the change in curvature from a normal distribution biased towards going straight again. I can't have the curvature go off the scale. Any off-the-shelf normal distribution functions?
That'll be one of the inputs, there'll be more inputs, outputs and weights but otherwise the structure is pretty much like the last program. If it gets near the edge of the road or turns the wheel too hard the car becomes increasingly unstable, so by the same principle as the last network it should learn the most predictable strategy. At some point I'll want to calculate the road curvature some way ahead of everything else so the NN can see the road ahead. I just realised that I can't acquire a huge list of snapshots because this thing might take a hell of a long time to learn. Maybe the printing hack was right: at least it saved memory. I have no use for the old values. I guess I should stick to the existing architecture but add stuff to the environment function that has nothing to do with the outputs but carries some state of its own along and feeds a subset of that state to the inputs. That makes this whole thread redundant. Sorry for the distraction. Adrian. ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
