Discovering Efficient Periodic Behaviours in Mechanical Systems via Neural Approximators