Extreme Adaptive Optics (AO) systems are designed to provide high resolution and high contrast observing capabilities on the largest ground-based telescopes through exquisite phase reconstruction accuracy. In that context, the pyramid wavefront sensor (P-WFS) has shown promise to deliver the means to provide such accuracy due to its high sensitivity. However, traditional methods cannot leverage the highly non-linear P-WFS measurements to their full potential. We present a predictive control method based on Reinforcement Learning (RL) for AO control with a P-WFS. The proposed approach is data-driven, has no assumptions about the system's evolution, and is non-linear due to the usage of neural networks. First, we discuss the challenges of using an RL control method with a P-WFS and propose solutions. Then, we show that our method outperforms an optimized integrator controller. Finally, we discuss its possible path for an actual implementation.