This paper introduces a multi-agent system (MAS) approach for modelling social behaviours in children?s play. The children, the dynamics of their play, and the playground are modelled into a MAS following the VOWELS [2] paradigm; thus, modelling the children as the agents, the playground as the environment, and the dynamics of their play as the interactions and organization. Utilizing the Playware playground [8, 9] as the environment, an interface to the environment can be defined, which allows us to formally model the children and the dynamics of their play (i.e. their individual and social behaviour) as populations and organizations in accordance with the PopOrg [3] paradigm. The MAS model has been used to design an artificial neural network (ANN) architecture capable of classifying social behaviours within appropriate categories. Experiments with children have been performed, and data recorded from the Playware playground during the experiments has been used to train the ANN. The ANN has been validated according to an external observer to correctly classify 70% of the social behaviours observed during the experiments.