next up previous contents
Next: OC1 Up: Inducers Previous: Aha Instance-based series

Perceptron and Winnow

Perceptron and Winnow are inducers that build linear discriminators. They are only capable of handling continuous attributes with no unknowns and two-class problems. For discrete data, you can use the ``conv'' utility to convert the input attributes to local encoding or binary encoding. The REMOVE_UNKNOWN_INST option can be used to remove instances with unknown values.

Perceptron uses the error correction rule (equation 5.19) in hertz-krogh-palmer. Winnow uses the algorithm described in littlestone-irrelevant.

All attributes are normalized to be in the range using extreme normalization (lowest values maps to 0, highest maps to 1). For different normalization types you can use cont-filter inducer as a preprocessor or run the ``conv'' utility. The reason for this normalization is that winnow overflows really fast when it raises numbers to powers.



Ronny Kohavi
Sun Oct 6 23:17:50 PDT 1996