I wanted to solve a regression problem, but I guess I do something wrong. I only have one output neuron what I would like to estimate me the prices of houses. Input neurons contain some input data what can affect the price of an actual house. So I encountered the following problems:
1) Weights do not change, they stay either -10, or 10. Should I somehow give an upper and lower bound? What if I don't know this bound because it might bea very big number. Then how to proceed?
2) I have boundary problem with the neurons as well. The same?
3) Is it possible to use more different kind of activation functions for regression problem such as RELU, RRELU, LRELU etc?
For me, this is completely not working, I am gladly sharing my data to whoever who might wanna check it.
Probably the first thing you should do is make sure you have rescaled all your data. You won't get good results if you have things like prices in hundreds of thousands to millions of dollars. So the inputs should be -1 to 1 or 0 to 1 generally. This should help with the weight issue but you can also just put very large (absolute) values in for upper/lower bounds. We don't have relu or its variants in Simbrain 3. You should be able to get backprop to do a basic regression for this kind of problem