Collective movements are pervasive behaviours among social organisms and have led to the development of many models. However, modelling animal trajectories and social interactions in simple bounded environments remains a challenge. Moreover, advances in the understanding of the sensory-motor loop and the information processing by animals are leading to revisions of the traditional assumptions made in decision-making algorithms. In this context, we develop a methodology based on artificial neural networks (ANN) to describe the collective motion of small zebrafish groups in a bounded environment. Although ANN models are commonly used in artificial systems they are still under-explored to model animal collective behaviours. Here, we present a methodology to calibrate Multilayer Perceptrons by learning from real fish experimental data. The ANNs are trained using either supervised learning or various forms of evolutionary reinforcement learning methods (using the CMA-ES and NSGA-III algorithms). We reveal that ANN models trained using evolutionary methods are capable of generating realistic collective motions for groups of 5 zebrafish including the tank wall effects, a feature that is lacking in previous models. Finally, we also discuss the benefits of optimised ANNs as candidates for driving robotic lure with biologically realistic behaviour, a method that is becoming increasingly popular to gather data and validate assumptions on collective behaviours.