The logic of reliable and efficient inquiry (Q1303795)

The author defends a learning theory approach to induction. He explores the consequences of three desiderata: An inductive method is \textit{reliable}, if it converges to the truth at some time \(n\). An inductive method is \textit{data minimal}, if there is no method that converges to the truth faster. An inductive method \textit{minimaxes retractions} if there is no other reliable method that needs a smaller number of mind changes on every datastream to converge to the truth. These are interesting notions, and they bear interesting relations to each other and to the topology of the hypothesis space. The basic ideas, definitions, and theorems are illustrated with reference to various complicated versions of Goodman's Riddle: Why should we prefer to say that all emeralds are \(green\), rather than that they are \(grue\), where \(grue\) is the property of being green until a fixed date \(d\), and blue thereafter. The author concludes (p. 426) that there is only one efficient inference rule for the infinitary version of Goodman's Riddle, and it is the natural rule: project that `all emeralds are green'. Gratifying as this may seem, it results from the structure of the hypothesis space. The author fails to take account of the importance of the \textit{fixed} date \(d\). The hypotheses in the infinitary case are to be expressed (in terms of the \(grue-bleen\) language) as \(Green(t)\): Emeralds examined before \(t\) are grue, but emeralds examined after \(t\) are bleen. Of course the uniquely efficient inference rule is to project the natural hypothesis, that `all emeralds are grue', as long as all observed emeralds \textit{are} grue. There is still, as Goodman observed, perfect symmetry.