Improving genetic algorithms for concept learning

Abstract

In this paper, we argue that the general learning abilities of genetic based techniques for concept learning can be improved in order to deal with numeric and symbolic values, tree-structured values, unknown values and user preference biases. The proposed algorithm, called SIA, uses the covering principle of AQ but with a genetic search that may be called several times. The genetic operators use a high level representation. The evaluation function can take into account the user preference biases about the learnt rules. The Iris database is taken as an example to underline SIA ability to deal with numeric values.