Fits a two-parameter logarithmic learning curve $$f(b) = c + a \log b$$ to the per-batch surrogate performance. Since the model is linear in \((c, a)\) the fit reduces to ordinary least squares on \((\log b, y)\) and standard errors come from the usual linear-regression covariance \(\hat\sigma^2 (X^\top X)^{-1}\) rather than from the Gauss-Newton delta method used by the nonlinear parametric LCE learners.
Training batches with batch_nr <= 0 are rejected because \(\log b\)
is undefined.
Creates a new instance of this learner.