Neural network training in Ruby

Question

I am new when it comes to neural networks. I fought ruby-fann and ai4r all day, and, unfortunately, I have nothing to show this, so I thought that I would come to Qaru and ask knowledgeable people.

I have a set of samples - every day has one data point, but they do not match the clear pattern that I could figure out (I tried a couple of regressions). However, I think it would be neat to see if there is any way to predict data going to the future only from a date, and I thought a neural network would be a good way to generate a function that could hope to express this connection.

Dates are DateTime objects, and data points are decimal numbers, such as 7.68. I converted DateTime objects to float, and then dividing by 10,000,000,000 to get a number from 0 to 1, and I divided the decimal numbers by 1000 to get a number from 0 to 1. I have over a thousand samples ... here's how Short exposure looks like:

[ ["2012-03-15", "7.68"], ["2012-03-14", "4.221"], ["2012-03-13", "12.212"], ["2012-03-12", "42.1"] ] 

What the conversion looks like:

 [ [0.13317696, 0.000768], [0.13316832, 0.0004221], [0.13315968, 0.0012212], [0.13315104, 0.00421] ] 

I would like this transformation not to be necessary, but I'm distracted. The problem is that both ai4r and ruby-fann return a single constant number, usually something in the middle of the range of patterns when I run them. Here is the code for ruby-fann:

 @fann = RubyFann::Standard.new(:num_inputs=>1, :hidden_neurons=>[3, 3], :num_outputs=>1) training_data = RubyFann::TrainData.new(:inputs => formatted_data.collect{|d| [d.first]}, :desired_outputs => formatted_data.collect{|d| [d.last]}) @fann.train_on_data(training_data, 1000, 1, 0.0001) @fann.run([DateTime.now.to_f / 10000000000.0]) # Always something random, and always the same number no matter what date I request it for 

And for ai4r:

 @ai4r = Ai4r::NeuralNetwork::Backpropagation.new([1, 3, 3, 1]) 1000.times do formatted_data.each do |data| @ai4r.train(data.first, data.last) end end @ai4r.eval([DateTime.now.to_f / 10000000000.0]) # A different result frmo above, but always seemingly random and the same for any requested date 

I feel like I'm missing something really basic here. I know this is a rather open-ended question, but if anyone can help me figure out how I am teaching my neural networks incorrectly, I would really appreciate it!

Answer 1

alfa has a good point in his comment, alternative ways to use NN may be more appropriate.

It depends on the problem, but if the value of the day even partially depends on the value of previous days, considering it as a time series, the Results can improve.

Instead, you would teach NN to produce the value of the day as a function of the window of, say, the previous ten-day values; you can also save the date parameter as a real input scale between [0, 1], as you think, it has a significant impact on the value of the day.