Talk:Negatively Correlated Ensemble Learning

REVIEWER A

This article gives the impression that NCL applies only to neural networks. It needs to be generalised.

NCL can be used without evolution. I would suggest to make the last paragraph of "Formulation of NCL" (except for its first sentence") part of a new section on evolutionary learning.

It would be nice to have some kind of conclusion.

The article should have more on NCL theory such as the ambiguity decomposition (see Diversity in Neural Network Ensembles by Gavin Brown).

=

REVIEWER B

Agreeing with the review above, this article seem overly biased toward (1) neural networks, (2) evolutionary algorithms, and (3) heuristic methods of choosing ensemble parameters, and especially (4) the author's own work. The NCL paradigm can be applied to any learning scheme that can minimize the specified error function, and is intrinsically *nothing* to do with evolutionary algorithms - an evolutionary algorithm can be used to optimize some parameters, but this is not *part* of the NC cost function.

There seems to be no mention of the strong links to well accepted ensemble theory, the bias-variance-covariance decomposition, and neither to results from "Managing Diversity in Regression Ensembles", (JMLR 2005) which explained both a bound on the strength parameter and evidence suggesting that the optimal value approached a constant with growing size of ensemble. This should be amended as it is a significant complement to the robustness of the NCL paradigm.

In addition other interesting pieces of theory have augmented NC, for example an interpretation as a Products of Experts in a PhD thesis by N. Edakunni, University of Edinburgh (chapter 2, section 2.2) ... a link to this is below:

 [[1]]

In general all the sections after the first one, i.e. "Constructive Training", "Ensemble Pruning", etc, seem to be extremely rushed, and without thought as to how they fit into the article overall. This should be addressed, and where necessary, more details included - being careful to keep the article focused on the NCL paradigm and how it has been developed.