@U0WL6FA77 Generally prefer wider rather than deeper networks. You probably can work with just 1 or 2 layers for many situations. Start with a narrow network (e.g. 5 neurons) and then try to over train your network. A good starting point is the same number of neurons as input variables. If the network can be over trained then your network is big enough (or perhaps too big). If your network is under trained then you can add more width (or depth) - but you will likely need to train for longer and with more data.

thank you What about next layers? The same number of units? half?

I would probably start with the second/output layer with the number of nodes = number of outputs. Then have the two layers fully connected.

BTW right now I have strange issue with keras when always have the same number in predictions whatever input is. In darts it works, but darts is too confusing to more detailed tests. Did you have such issue before to always have the same number in prediction? Is it possible? Or is it just some kind of strange bug? I debugged input arrays and there are different values for sure.

I haven't encountered this issue I'm afraid. No idea is that would suggest a bug in the library or in your usage of it. If you are not confident in your usage code then I would try to train a "hello world" network first. e.g. 1 input, 1 output, 1 layer. Overtrain the network and make sure it can predict correctly your training data. • 0.1-> 0 • 0.2 -> 0 • 0.3 -> 0 • 0.4 -> 1 • 0.5 -> 1 • ...