A conceptual approach to study spatial movements of fish using an individual-based neural network genetic algorithm model is presented. Artificial neural networks, where the weights are adapted using a genetic algorithm, are applied to evolve individual movement behaviour in a spatially heterogeneous and seasonal environment. A 2D physical model (for the Barents Sea) creates monthly temperature fields, which again are used to calculate zooplankton production and predation pressure. Daily fish movement is controlled by reactive or predictive mechanisms. Reactive movement governs search for local optimal habitats, whereas predictive control enables adaptation to seasonal changes. Levels of growth and predation pressure at the time of decision are used to assess whether to apply reactive or predictive movement control. To make the model realistic on a large scale, each of the individuals are scaled up to represent a clone of one million siblings acting and growing synchronously. The fish lives for up to two years, and may reproduce in its second year. In order to spawn it has to be at the designated spawning area in the south-western part of the lattice in January. During spawning it produces a number of offspring in proportion to its body size. The "genetic constitution" of offspring (the weights of the synapses in the neural networks) is a mix of their "mother's" and a randomly picked member of the population. The model is able to solve the problem of navigating in a heterogeneous and seasonal environment. The movement of the artificial fish follows a seasonal pattern, typical for migrating pelagic fish stocks. During summer and autumn the distribution is widespread whereas during spring it is more concentrated. When trophic feedback is removed (i.e. zooplankton survival is independent of fish predation) the distribution of the fish is less dispersed which shows that the model allows for density dependent behaviour. Large-scale migration is an interplay between reactive and predictive movement control and when only one of these is allowed, the individuals are unable to adapt properly. Throughout most of its life the fish relies heavily on reactive movement, but during the spawning migration predictive movement control is applied almost exclusively. (C) 1998 Elsevier Science B.V. All rights reserved.
