Scientists have a strong desire to catch real life in models. But which model is the best? Steven de Rooij from the Centrum Wiskunde & Informatica (CWI) in Amsterdam researched model selection with computational techniques in his thesis Minimum Description Length Model Selection - Problems and Extensions. On 10 September he receives his PhD at the University of Amsterdam. The results could be applied to the prediction of share prices, filtering of unwanted email, data compression, economic models, and climate change research.
A model can be seen as a theory, or concise description of reality. It can represent a scientific hypothesis (‘E = ½ mv²'), a text comprehending a speech recognition program (the user said: ‘chair') or impressions from the outside world ("I think he/she likes me!"). The issue with such models is that many alternatives are at hand (in this case for instance: ‘E = mc²', ‘fair' and ‘jerk'). It is not immediately clear how to choose the best model.
When models are specified exactly, based on observations, the Minimum Description Length principal is a good method for selection. Until now it was not taken into account, however, that which model is most suitable changes in time. This is due to changing circumstances, or the fact that models start to behave differently when there are more observations. De Rooij found a way to make existing models dynamic. This way the best model can be selected in spite of changing circumstances.
