Algorithmic statistics, prediction and machine learning
This work addresses foundational issues in algorithmic information theory for researchers in theoretical computer science and machine learning, but it appears incremental as it builds on existing frameworks.
The paper tackles the problem of extending algorithmic statistics to include prediction and machine learning contexts, showing that a hierarchy based on algorithmic statistics and a priori probability are equivalent for prediction, and extending the framework to handle sets of positive examples in learning.
Algorithmic statistics considers the following problem: given a binary string $x$ (e.g., some experimental data), find a "good" explanation of this data. It uses algorithmic information theory to define formally what is a good explanation. In this paper we extend this framework in two directions. First, the explanations are not only interesting in themselves but also used for prediction: we want to know what kind of data we may reasonably expect in similar situations (repeating the same experiment). We show that some kind of hierarchy can be constructed both in terms of algorithmic statistics and using the notion of a priori probability, and these two approaches turn out to be equivalent. Second, a more realistic approach that goes back to machine learning theory, assumes that we have not a single data string $x$ but some set of "positive examples" $x_1,\ldots,x_l$ that all belong to some unknown set $A$, a property that we want to learn. We want this set $A$ to contain all positive examples and to be as small and simple as possible. We show how algorithmic statistic can be extended to cover this situation.