Can artificial neural networks simulate fish movement in novel environments?

Spatially-explicit population and community models are increasingly being used to address issues in fisheries and conservation. Accurately simulating behavioral movement is critical in many of these models. Artificial neural networks (ANNs) trained with genetic algorithms (GAs) have been used to simulate vertical and  spawning migrations. We evaluated the ability of ANN movement rules to respond to novel gradients with a simple, individual-based model. Predation pressure was distributed in different spatial patterns on the grid, and individuals attempted to minimize their cumulative predation mortality. Mortality, lagged mortality, and swimming angle were used as inputs to the ANNs. We first demonstrated that the ANNs could be trained using back-propagation with a training data set generated by moving individuals using a neighborhood search algorithm. The ANNs were successfully trained (individuals moved toward the minimum predation area) with back-propagation for three mortality grids. We then showed that the GA could achieve similar results. However, neither method estimated ANNs that could predict movement on a novel mortality grid. Thus, care should be taken when using ANNs to include all anticipated gradients during the training phase. Next steps should focus on more sophisticated ANNs and GAs that can predict movement on novel gradients with multiple cues.