First, generate a data-set by clicking on the left and right buttons in the main window of the applet. Then, press "split" to split the data into training and test sets. Then you will be able to press "step" for a single boosting iteration, or "10 steps" for executing 10 steps at a time (most of the calculation time of the applet is spent updating the graphics, so in this way you can save a lot of time).

you can press "reset" to reset the boosting process. Then you can either re-split the data, or press "edit" to go back and update the data. In edit mode you can press "clear" to start from an empty set of examples.

Using the check-boxes on the top you can control to display in the main window. The left selectors choose whether the image is the decision rule (the default) or the sum of the weak hypotheses before thresholding.

The right checkboxes let you display all the data, or just the training set or just the test set. The graphs below the main window show the history of the errors as a function of the iteration number.

• blue = training error
• green = test error
• yellow = error of the weak hypothesis (on the weighted training sample).
• red = theoretical bound on the training error.

A bug: sometimes the bound on the training error gets below the actual training error, maybe there are rounding problems?

todo:

1. Display the current error values in the applet in the color that corresponds to the graph colors.
2. load datasets
3. other weak learners: tree-based. non-axis-parallel planes.