hauenste@lan.informatik.tu-muenchen.dbp.de (Andreas Hauenstein) (02/07/90)
For my thesis, I need to implement adaptive classifiers. A classifier is a system that is given feature vectors (x1, ... , xn) and, with each feature vector, a class it belongs to. From these examples, it is supposed to learn to assign feature vectors to classes. In my case, the vectors (x1, ..., xn) consist of numbers (integer or float), not of symbolic expressions. Symbolic approaches are not what I'm looking for. There is a lot of literature about the topic, e.g. Duda/Hart, but there is _ALWAYS_ a strong distinction between training phase (when the examples are given) and application phase. After the system has been trained on the training data, it is supposed to be static forever. What I am looking for is an ADAPTIVE classifier. The training phase and the application phase should not be distinct, and the classifier should be able to change with time. It should not just improve by refining class boundaries, like Quinlan's ID3 does, it should even be able to change the class areas completely (over time), if a change in the environment should make it necessary. Class boundaries can be simple, e.g. bricks will do (i.e. rectangular right-angled prisms). Of course, it should be cheap and fast :-). Neural nets are two slow and not known well enough, I think :-(. I'd be glad to learn that I'm wrong, though. If anyone knows of even moderately successful implementations of such systems, or about literature on the topic, PLEASE tell me about it. Thanks, Andreas.