1
$\begingroup$

The inputs (features) and expected output for my ANN are these:

  • Input 1: Product id (number, cast to double)
  • Input 2: Year in the past (1900..2017, cast to double)
  • Input 3: Month of year (1..12, cast to double)
  • Expected output: Sale of month (number of units sold, cast to double)

I need to predict the sale of a product for a certain month in a certain year. How many layers and how many neurons on there layers should I put?

$\endgroup$
  • 1
    $\begingroup$ Figure it out through trial and error. You don't have a lot of inputs, so you should not make it to big. I think having two hidden layers is enough, with a maximum of 10 neurons per hidden layer. But what have you tried? Because you have such a small input/output it would be fairly easy to figure it out yourself through trial and error... $\endgroup$ – Thomas W May 9 '17 at 9:33
  • $\begingroup$ i have 12 training samples for the 12 months of last year * 1 product. I use this shape [10,10,1], but the training doesn't converge after 30,000,000 sample iterations $\endgroup$ – datdinhquoc May 9 '17 at 9:43
  • 1
    $\begingroup$ Have you tried using an LSTM? $\endgroup$ – Thomas W May 9 '17 at 9:48
  • $\begingroup$ never tried LSTM before, may be because of something wrong in my own ANN code, i'll try some libraries $\endgroup$ – datdinhquoc May 9 '17 at 9:58
1
$\begingroup$

Every answer you get is just an opinion, it is based on experience. My answer is a single hidden layer of 5 neurons.

Besides, I recommend you to use a TWEANN system (Topology and Weight Evolving ANN).

Such system applys genetic algorithm to search and optimize an ANN with the optimal topology and weights.

Take a look at NEAT and DXNN.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.