Recently I've started playing with neural networks (I use brain.js library) and I have a problem with training my network. I want to create a neural network which would give a certain number (in this situation probability) as an output. The input data comes from sensor which detects a certain event. I provide data in the following format:

{ input: "1 4 0",
  output: "0.003"}

The numbers in input corrensponds to: index of time range when the sensor made a measurement, day of week when the measurement was done and the last one is measurement itsefl (number 0 or 1). They are in text, because when I tried to learn neural network with numbers, the output were even worse. I train network with around 3300 sensor's measurements. I tried to learn the network with much smaller amount of data (~30) and it took around 20-30min to train and give output which was a good output.

I tried to use LSTM neural network with following configuration:

 const config = {
   binaryThresh: 0.5,
   hiddenLayers: [3], 
   activation: 'relu', 
 };

and training:

const trainConfig = {
  iterations: 20000, 
  errorThresh: 0.01, 
  log: true, 
  logPeriod: 1000, 
  learningRate: 0.5, 
  momentum: 0.1, 
  callback: null, 
  callbackPeriod: 10, 
  timeout: 600000, 
}

No matter how much I change this configuration the training starts with enormous training Error which ends up on 'Infinity' but despite that training is still running.

Could someone help me and tell me what I'm doing wrong? I tried to change even format of data :(