Abstract
Growing concerns over sustainability between human activity, natural resources and ecosystems have been widely recognized in the world. Due to the complex systems and processes involved, it is quite difficult to quantify and find an optimal strategy for sustainable development and management. Artificial neural networks (ANNs) have widely been applied to modelling complex systems such as the relationship between habitat conditions and the distribution of target species. Since such relationships are strongly affected by the uncertainties originating from errors in data and modelling processes, fuzzy neural networks (FNNs), a fuzzified version of ANNs, have been applied to habitat prediction. Despite the high predictive ability of FNNs, they suffer from problems such as variances in model structure and overfitting to given data. The present study, therefore, aims to evaluate the effect of a weight decay backpropagation on the accuracy and model structure of the FNNs that evaluate the spatial distribution and habitat preference of Japanese medaka (Oryzias latipes). As a result, weight decay improved the generalization ability of the FNNs with a small deterioration of predictive accuracy. The habitat preference curves derived from the FNNs could represent the trends of habitat preference of the fish, of which the variance was markedly reduced. This case study of Japanese medaka supports the applicability of FNNs with weight decay, which can retrieve consistent information from the field observation data.
