Results (
Thai) 1:
[Copy]Copied!
The desired output is of course the 3x3 identity matrix, matching the bipolar target matrix in the routine. The results show that the network has not fully learned the noiseless patterns after 20 and 30 epochs, but has achieved success after 40 iterations. After learning is complete the net is able to correctly classify the noisy signals for signal to noise ratios of 100 and 10, but has problems when this ratio is 5 or less.Note that even for very large noise values the first row of the output matrix is correct in each case. This result is not unexpected since the square wave differs considerably from the other two waveforms and can therefore more easily be classified even with large noise superimposed. On the other hand, the triangular and sine waves are closer in appearance and are therefore more difficult to distinguish. The similarity is evident in Figures 5(b) and 5(c) with noise superimposed.Note also that when the program is run a second time with the signal to noise ratio set to 1, different outputs result due to the randomness of the noise. Initializing the random number seed was not done in the program.D. Signal Frequency Separation using Perceptron A modified MATLAB signal classification program trains a neural network to classify three sinusoidal signals of the same amplitude and phase, and separated only in frequency. The middle frequency is Δ percent above and the highest frequency 2Δ percent above the lowest frequency. As for the previous example, after training an attempt is made to associate noisy signals with the learned signals. The parametersentered upon running the program are the percent frequency separation, the number of samples per period, the number of training epochs and the signal to noise ratio.
Being translated, please wait..
