The quadruped locomotion task is generally conditioned on a velocity command that is user-defined. However, certain features of the locomotion are not represented in this high-level definition of the task, e.g., swing feet height, step length and expended energy. Using reinforcement learning, many of these features can be determined by the reward function terms and scales that are often fixed. In this work, we propose a deep reinforcement learning (DRL) approach to learn control policies augmented with parameters that modify different aspects of the reward function and control setup which, in turn, result in variations of the locomotion. We can then define a hierarchical architecture where a high level policy infers the suitable parameters to complete a given task. We show that this setup makes it possible to learn more complex behaviours that can be adapted for different terrains and environments to ensure successful and efficient locomotion. We display our results on the MIT Mini-Cheetah quadruped.