Recent advances in media-related technologies, including capturing and processing, have facilitated novel forms of 3D media content, increasing the degree of user immersion. In order to ensure these technologies can readily support the rising demand for more captivating entertainment, both the production and delivery mechanisms should be transformed to support the application of media or network-related optimizations and refinements on-the-fly. Network peculiarities deriving from geographic and other factors make it difficult for a greedy or a supervised machine learning algorithm to successfully foresee the need for reconfiguration of the content production or delivery procedures. For these reasons, Reinforcement Learning (RL) approaches have lately gained popularity as partial information on the environment is enough for an algorithm to begin its training and converge to an optimal policy. The contribution of this work is a Cognitive Network Optimizer (CNO) in the form of an RL agent, designed to perform corrective actions on both the production and consumption ends of an immersive 3D media platform, depending on a collection of real-time monitoring parameters, including infrastructure, application-level and quality of experience (QoE) metrics. Our work demonstrates CNO approaches with different foci, i.e., a greedy maximization of the users' QoE, a QoE-focused RL approach and a combined QoE-and-Cost RL approach.