JOSÉ HERNÁNDEZ-ORALLO and ISMAEL GARCÍA-VAREA EXPLANATORY AND CREATIVE ALTERNATIVES TO THE MDL PRINCIPLE

The Minimum Description Length (MDL) principle is the modern formalisation of Occam’s razor. It has been extensively and successfully used in machine learning (ML), especially for noisy and long sources of data. However, the MDL principle presents some paradoxes and inconveniences. After discussing all these, we address two of the most relevant: lack of explanation and lack of creativity. We present new alternatives to address these problems. The first one, intensional complexity, avoids extensional parts in a description, so distributing compression ratio in a more even way than the MDL principle. The second one, information gain, forces that the hypothesis is informative (or computationally hard to discover) wrt. the evidence, so giving a formal definition of what is to discover.