The purpose of this demonstration is to help you develop intuition for how support-vector learning performs. The particular problem solved is that of separating the population of positive samples from the population of negative samples.
The Kernel form menu item on the menu bar provides access to either the dot-product or radial-basis kernel functions. The Test library item enables you to select from several sample arrangements of positive and negative samples.
You press the Learn button to initiate learning. The Clear button gets rid of all samples. The next button down from the Clear button varies the exponent if you are using a dot product kernel; the same button varies the width of the kernel if you are using a radial-basis kernel. The Refresh button refreshes the display, which is sometimes helpful if the display program gets stuck.
The main display shows the values returned by the support vector machine. Gray means plus or minus one. More blue means more positive than one; more red means more negative than minus one.
Under the main display, you see how optimization has progressed. It changes each time you learn, even if you haven't changed the samples, because there is a random element that prevents particular sample-pair overuse in the optimization process.
You also see a display of the support-vector weights. In most cases, most will be zero. They do not change much, even if you relearn, because the learning process finds a global maximum. Any change you see is a result of the learning process coming within tolerance by different routes, as determined by the random element that prevents sample-pair overuse.
Support Vector Java® Demonstration
Cick on the Java® link and Open the file when prompted.
If the demonstration does not appear, you are probably using a web browser that does not support the version of Java® in which demonstration was written, so you will need a so-called plug in. If your browser does not volunteer to load the plug in for you, you may need to load the plug in from the Sun plug-in website.
Testing was done with Java® 1.4, running in Netscape Navigator 7.02 and in Internet Explorer 6.0.
Java® is a trademark of Sun Microsystems, Inc. in the United States and other countries.