Improved estimation, using neural networks, of the food consumption of fish populations

The aim of the present work is to improve the relevance of methods to predict the Q/B ratio (annual consumption of food Q relative to the biomass B of fish species), which is essential for any multispecies stock model based on trophic relationships. Two methods were considered: multiple linear regression (MLR), improved by the log transformation of some variables, and artificial neural networks (NNs), which have the advantage of accepting nonlinearity in the relations between Q/B and different independent variables. Although MLR is acceptable for predicting small values of Q/B (mainly carnivorous fish), it does not display good performances for high values (herbivorous and detritivorous fish). In contrast, by using the gradient back-propagation algorithm, the NNs are suitable for a valid estimation of the whole range of known values of Q/B. Both types of model were tested with test sets of data (drawn at random from the full set of data) that had not been used for model construction. The proposed methods are thus predictive. As they require only a few easily accessible parameters, they can avoid tedious studies of fish feeding over a daily and an annual cycle. The NN program used, operating on a personal computer, is available on request.