Learning CPG-based Biped Locomotion with a Policy Gradient Method: Application to a Humanoid Robot | Synapse