So I am struggling to wrap my head around this so I have been messing around with some very simple neural networks with at most using 2 layers. Basic things where the network will generate a rule (for example an equation of a line) based on guessing the result of an event.
I am wanting to step this up to giving my network for inputs:
An array of objects with their radius and for outputs:
An x and y direction vector.
If the ship hits the object it will be given an error.
So here is my question,
What kind of qualitative error can I give the neural network to learn since I'm not sure if just saying 0 is not crashing and 1 is crashing will produce effective results.
Secondly, is this kind of thing possible using just perception, weighting with biased. If so how many perceptions will I be need and how will the be arranged.
I may be missing something fundamental here I just don't understand how a bunch of weightings on variables can allow a machine to navigate around an object.
This is me trying in preparation for my 3rd year report on neural networking.
