The role of relevance in explanation. I: Irrelevance as statistical independence (Q2366556)

In a typical case when we do not have complete information about the world, several explanations are possible for every observable fact. Which of them should we choose? A seemingly natural idea is to choose the most probable explanation. This idea, however, is not flawless: suppose, e.g., that a pavement is wet, and that there are two possible explanations: that it has been raining at night (probability 99.9\%), and that some kid soaked the pavement with a water gun (probability 0.1\%). Surely rain is the most reasonable explanation. However, since we decided to take weather into consideration, we may want to consider ``rain'' not as one alternative, but as several (like ``rain and wind 2.5 mph''). If we thus subdivide the rain case into 1000 different subcases, then for each of them probability is \(<0.1\%\), so a kid becomes the most probable explanation. How to avoid this absurd conclusion? The author proposes to restrict alterantives to only those that are relevant (= not statistically independent) with the observed facts (in our case, direction and the strength of the wind has nothing to do with the fact of the pavement being wet or, not, so these parameters are completely irrelevant). The author shows that in general, finding the most probable explanation is NP-hard; he also provides an algorithm that works reasonably well in realistic situations.