Section: Research Program

Incremental learning

The first learning algorithms were batch learning. They examine all examples and produce a concept description, that is generally not further modified. This is not adapted to dynamic settings where data are delivered continuously. For such settings, incremental algorithms have been proposed. These algorithms examine the training example one at a time (or set by set), maintaining a "best-so-far" description which may be modified each time a new example (or set of examples) arrives. In order to strengthen the learning process, some specific old examples are often kept: this is called partial memory systems. A more specific classification of incremental learning can be found in [74] .

Current issues in incremental learning are

• for partial instance memory: how to select examples, [72]

• the problem of hidden: the target concept may depend on unknown variables, which are not given as explicit attributes [84]

• the problem of concept drift: the target changes with time [83] , [57]

• the problem of masked example: the data distribution may change and some examples may not be anymore visible.

In many application domains, model inference and further modifications have to be validated by an expert. Thus, the model should be a glass box and its representation language should be easily understandable by a human expert. This is why we investigate rule-based formalisms for incremental learning [57] .