I am building a small tool to train neural networks for Unity games using genetic algorithms.
I want my framework to be generic and usable with multiple games. One very important part regarding performance and training time is the normalization of inputs.
Depending on the game, we may want to normalize inputs across different dimensions, to normalize with respect to a maximum theoretical value ... or not normalize at all.
For example, if the three inputs of my neural net are a ball position (x,y,z), I will normalize each input regarding its maximum theoretical value (assuming I know it).
My question is:
- Is the best practice to let the normalization process be part of the game development, or to implement it in my machine learning framework ?
- If I decide to implement it in my framework, what is a good tool to build flexible normalization on inputs, that the user could configure according its game ?
