76561198012850962 if you’ll get some good tyre machines out of it, please send them to me if you don’t mind, so CM would have something by default.
One other question, please: FANN is said to use the RPROP algorithm as default.
I tried various approaches and found this set of parameters to work better, while RPROP produces some weird results:
(And it’s a result which were averaged between four networks.)
It could be because I need to adjust training runs or layers, but I think I did it to no positive effect. My suspicion is that RPROP was designed to handle more data (I think 60 training samples usually is way too little data for neural networks, but it’s all we got) and networks more complex than that.
I don’t really know too much about how it all works, but if you do, by the way, can you please tell me if it’s better to use separate networks or a single network with ≈50 outputs? My thinking is that, on one hand, separate networks are independent from each other, so there’ll be less senseless noise. But, on the other hand, maybe single networks having all the data could detect patterns more reliably? There is an option, but with a single network, I didn’t manage to get any good results.