Rosenblatt's Perceptron Learning Algorithm

Below is a Java program which implements Rosenblatt's Perceptron Learning Algorithm (PLA). The PLA is a method which can determine a weight vector which is normal to the plane which divides vectors of two classes if the two classes are linearly separable. In simple terms, if there are two clusters of points of different "types", then this algorithm can learn to find the line which separates the two clusters.


Instructions



Comments

The threshold step size affects the rate at which the third component of the weight vector, the threshold, is updated. A value of 25.0 is good in most cases, though larger values can speed learning if the separating line is far from the origin, and slow it down if the line is closer. This was added just because the other components of the input vector were much larger than 1.

The learning rate is the traditional learning rate which scales the training vector before being added to the weight vector during updates. Try different values if you like, larger values will affect the convergence rate, but not the convergence itself. (i.e. pick a large value and it'll still converge).

The number of iterations is put in so that the algorithm won't run forever if the input vectors are non-separable. Keep clicking on "Go!" until the status line says that learning is complete, or you are convinced the input vectors are non-separable.


NOTE: If you do not see anything above, make sure you have Java enabled in your browser. For Netscape that means, click on Options in the menubar, and select Network Preferences. Then click on the Languages tab and from there click on Enable Java. Then hit Ok and then click on Reload. You might want to save your options for the next time you run Netscape (Options->Save Options).


Last updated: Tue Sep 10 17:27:23 EDT 1996, jesse@mit.edu
© 1996 - W. Hong, Massachusetts Institute of Technology
All Rights Reserved