Portal de Periódicos da CAPES

Discussion of “Generalized regression neural networks for evapotranspiration modelling”

2007; Taylor & Francis; Volume: 52; Issue: 4 Linguagem: Inglês

10.1623/hysj.52.4.832

2150-3435

Demetris Koutsoyiannis,

Hydrology and Watershed Management Studies

There is no doubt that so-called “artificial neural networks” (ANN) are powerful computational tools to model complex nonlinear systems. In my view, an ANN establishes a data-driven nonlinear relationship between inputs and outputs of a system. The fact that such a nonlinear model is generally very complicated (usually one does not even write down the equations) renders it a black-box model. The fact that the model contains numerous parameters makes imperative the use of an advanced optimization method to calibrate its parameters. Once an ANN is fitted, it can be used to predict outputs from known inputs. Thus, there have been numerous successful applications of ANN in forecasting the future evolution of complex systems (e.g. Casdagli & Eubank, 1992; Weigend & Gershenfeld, 1994). However, I am afraid that there has also been an abuse in other cases, indirectly assisted by the numerous technical details, inapproachable for the majority of scientists (in our case hydrologists), and even by the exotic ANN vocabulary. The paper by Kisi (2006) stimulated my “reflex” questions about “neural networks”, their use and abuse, and helped me to organize them so that they can be addressed to the “central nervous system”.