Fast reinforcement learning through the composition of behaviours