The MML Evolution of Classi cation

Minimum encoding induction (MML and MDL) is well developed theoretically and is currently being employed in two central areas of investigation in machine learning|namely, classiication learning and the learning of causal networks. MML and MDL ooer important tools for the evaluation of models, but ooer little direct help in the problem of how to conduct the search through the model space. Here we combine MML with one of the more powerful search techniques available to machine learning: genetic algorithms (GAs). We develop a genetic algorithm to search the space of classiication (decision) graphs using an MML-based tness criterion and establish its eeectiveness across a range of test cases from the UC Irvine machine learning archive.