How we train our artificial neural networks.
Artificial neural networks ( ANN ) in general is a modern mathematical tool which is used in a great variety of practical applications
The main advantage of this tool that it can be trained as a human brain.
There are several methods of training of ANN.
We use so called supervised training.
For training purposes we collected as of today 19 thousand love stories with known outcome.
Approximately 80 percents of these stories were used for training.
And 20 percent for checking the training quality.
To better understand our training process please look on the picture.
We will try to explain with very simple words using no special math .
Training process goes step by step.
First step - we take data of the first pair of people (for every person - blood type, color of eyes, height, date of birth, time of birth, geographical coordinates of birth, timezone etc) and use them as an input to our network.
The results of their relationships are on the output. The results for first three rows of output are calibrated by degrees - very bad, bad, neutral, good, very good and additionally by scale of strength from 1 to 100. The fourth row in output is a number of years people spent together before separation.
Second step - we take data of next pair of people, use it as an input. As output we use the results of relationships for this pair.
And so on.
At the same time we make necessary training mathematical steps on our network.
Artificial neural network has an ability to recognise function between input and output information after a certain amount of information which was given to them. The more information we give for training - the better is the understanding. It is like a human brain.
Also artificial neural networks have an ability to say, that this or that input of information has no connection with the results. This means that this specific information on input is not influencing on the result.
In our case at this moment we got a result that blood type, color of eyes, height are not influencing on the result of relationships.
But the precise birth data (with birth time) and birthplace coordinates are influencing.
After finishing training of our network on 80 percent of collected data we conduct a check how well the network is trained using the remaining 20 percent of data. For this we gave step by step input data of remaining pairs as an input and asked the network on its opinion about output of this union and compared the real output with the predicted output.
If the predicted output and the real result of relationships are close to each other in majority of cases we say that this network is trained well and that this network can predict the output on any other given set of input information with sufficient precision.