Could a solution be to convert my input/training data to the scale of 0 to 1?
At the moment, my inputs are relative to what they are on the screen, eg. bottom-left corner point is [-0.82, -0.82], would converting that to 0.2 or whatever be of a help?
Right, after discussing it with my project supervisor today, my problem appears to be that when I'm training the network, I do not have any projected outputs.
The original network featured binary from '000' to '111', with an XOR 4th field for the projected outcome. I know that for my example I'm going to need to have 3 outputs (rotation, speed of X and speed of Y).
I know how to rememdy this (just add extra values to the training data), but once again I am stuck at the concept of trying get out these projected results from the network.
Thanks for your suggestions thus far and any further help will be GREATLY appreciated.
Only some 3 hours before deadline (with testing, evaluation and completing the report plus printing and binding), the lecturer who made the network suggested just having 3 separate networks for each of the three outputs.
Grrrr. I thought of that over a week ago, but didn't go through with it due to the fact that I thought the purpose would be to have 1 neural network. In fact, my whole damn report was based around that. So cue frantic coding, poor testing and write-up and I've finished. Not what I needed some 3 hours before deadline, having been awake for approx. 26 hours at that point, pulling an all-nighter.
Whilst the car didn't travel exactly perfectly (it moves off the screen or in strange positions) it is clearly much better than I had previously, and only spending more time with test and training data and whatnot would improve it IMO.
If anyone would like to check out my fine result I'll gladly upload it somewhere, in the meantime, thanks for the help anyway.
