neural net on NBA derivatives

[339955]

i am modeling a neural net for 2nd half totals with 6 inputs (fg total, fg spread, favorite 1st half score, dog 1st half score, how much favorite is beating 1st half spread by, how much game went over 1st half total), with 15 hidden units, and then the output. how come my NN never converges and the output it gives varies widely from one training session to another? how do i fix this?

[Miz]

What type of network did you choose? Did you try varying the learning rate?

[cyrusii]

Using neural nets is an art. You really need to study neural nets, be able to understand how they really work, and have experience with using the correct form of nns. Most people look at them as some sort of black box, that would never work. I suggest reading a book. Neural nets are getting obsolete anyways.

I used to be a TA for neural net course in college.

[Miz]

Using neural nets is an art. You really need to study neural nets, be able to understand how they really work, and have experience with using the correct form of nns. Most people look at them as some sort of black box, that would never work. I suggest reading a book. Neural nets are getting obsolete anyways.

I used to be a TA for neural net course in college.

I agree. I'm not a NN expert, but most of the time I saw great variation in performance, and it is difficult to know what to begin with. How many hidden layers etc. I also think they seem to work best as classifiers, not quantifiers. In other words if you have 20 different qualitative inputs and the output is a color... then the NN is good at determining the most likely color for a new test set of 20 different inputs. Beyond that they are of limited value for sports betting. There is other work being done that is far more valuable, but I'll keep that to myself.

[339955]

Using neural nets is an art. You really need to study neural nets, be able to understand how they really work, and have experience with using the correct form of nns. Most people look at them as some sort of black box, that would never work. I suggest reading a book. Neural nets are getting obsolete anyways.

I used to be a TA for neural net course in college.

thanks for your advice cyrusii. what type of analysis is making NN's obsolete?