#mainTextBox We have now switched to trying to predict a square wave. The discontinuity in the signal will be harder to predict than the sine wave. We have a 32 sample square wave and a 15-tap delay line. Run the network. The network cannot accurately predict the discontinuity. Change the desired and input signal generators to 30 samples per cycle. Now run the network. It predicted perfectly. How can a linear network predict this nonlinear discontinuity? Look at the weights. Now do you understand? The delay line is long enough to see both the current and last transition -- thus, it can look for 15 consecutive values and then predict a change. The weights reflect this. All weights are near zero except the first and 15th. These two points are the only thing needed to correctly predict the signal.
#subtitleTextBox Summary
#mainTextBox Prediction is just a special case of filtering where the desired signal is the input signal one step ahead. Try different signals and different tap delay lengths. Look at the weights and the spectrum of the system to understand what is going on.
