Neural network / maths assistance?

Hi all,

Thank you for taking the time to read my thread - I'm sorry if I have posted it in the wrong place.

I am in the process of making a type of neural network out of a 'labour of love', to analyse some data that I have accumulated over the past few years. Sadly, I'm not a computer scientist, but I do understand a reasonable portion of the maths and the programming. I'm getting stuck on a few maths/data scaling issues where I could really use the guidance of a subject-matter expert, to make sure I am on the right track.

I know its a slight odd question to ask on an Excel forum, but wondered if anyone knows of any good forums where maths / neural networks / data-pre-processing, etc are discussed, as I dont think the ExcelForum is really intended for that!

Thank you for any assistance that anyone can offer.

Paul

Hi guys,

Thanks for taking the time to reply to me. I am sorry for the delayed response.

I am trying to format a range of data so that it is suitable for input to a neural network (that I have yet to make, so getting the input data in an appropriate format is the first step!). I have read many articles, and they state that if you have two inputs that are different scales, they should be adjusted so that they are more comparable. For example:

InputA has a range of 0 to 1
InputB has a range of 1 to 500

The articles say that this (to a neural network) would make it look like the InputB data is 500 times more important/bigger, and whilst the network weights will eventually adjust to compensate, its hardly giving the network the best possible start.

I have seen a few formulae for rescaling the data, for example that would make the InputB data range from 0-1. My problem is that in my particular case, for the InputB data there are more datapoints around the 20-50 mark, and the rest is more sporadic. The trouble is that there isnt much I can "do" about this distribution of the input data - thats just how it is, but im worried that it will affect my network or results somehow, or when I rescale the data to (say) 0-1, the values that used to be in the 20-50 mark will effectively become the same value because of the way they are scaled?

Does this make any sense? I am not sure I have explained it particularly well!

My other issue is about the network output. I am presenting put data with its corresponding "target" output (i.e. the desired result) so that the network can be trained. My problem is that the neural network is looking at betting data, and I am having trouble understanding how to specify the results.

I originally thought that I would have 0 for a loss, and 1 for a win, because its nice and simple. But the trouble is that betting isnt that simple, because not all selections have the same odds. The system could be forgiven for selecting a 50-to-1 bet that loses, but shouldnt be picking 1-to-5 bets that lose, because obviously they're "expected" to win. I considered doing a scoring system, where for a win the output target scope is 1 * (the odds), so that bigger wins have a bigger score, and for losers the score would be "output = -1 / odds". This means that a big odds loser scores a low (negative) score, whereas a short odds loser (which is worse) will score a more negative score.

I think that the problem with the above is that the "output function" is not continuous, and therefore might confuse the network and cause me to get "duff" results.

Has anyone had much practice with formatting inputs / outputs to neural networks and if the above makes any sense, would it be possible for them to share their knowledge?

Thanks again for the responses and any help you can offer.

Paul