Learning theorems (Q416397)

Let \(x_{0}, x_{1},\dots\) be an infinite sequence of points of metric space \(M\) and an unknown real valued function \(f: M \rightarrow M\), suppose that we have learned a sequence of \(n\) data points \((x_{0}, f(x_{0})),\dots,(x_{n-1}, f(x_{n-1}))\). In the introduction the author writes: How to predict the value \(f(x_{n})\)? Assuming some regularity of \(f\) the simplest way that appears reasonable is the following. Define a function \(f_{n}: M\rightarrow \mathbb{R}\) (\(n > 0\)) such that for every \(x \in M\) we pick the last term \(x_{k}\) of the sequence \(x_{0},\dots, x_{n-1}\) among those which are the nearest to \(x\) and let \(f_{n}(x)=f(x_{k})\). Moreover, the author pointed out: The main purpose of this paper is to prove other theorems of that kind related to the Laws of Large Numbers. We add also a little improvement of a convergence theorem of the Kaczmarz-Agmon Projection Algorithm. As a complement to signal the results contained in this paper one can quote the abstract: We will prove learning theorems that could explain, if only a little, how some organisms generalize information that they get from their senses.