Backprop

MLPs are cool, but what makes them popular are learning functions that make them able to learn to approximate a function represented by data.
There are different rules, but the original and most widely used is the backpropagation of error, or simply backprop.
The backprop wiki has a full explanation, but here's a synopsis.
Remember each input and output from a perceptron is weighted by the connection
The algorithm changes these weights to better approximate the data.
To do so, run the MLP and get the output. It's a supervised algorithm so it needs to know the answer or answers.
Using the difference between the predicted and the actual answer, change the weights. Do that for the next layer (and then again for the next layers if they exist).
Keep running the algorithm until its right on all answers or you get tired.