Exemplary embodiments can maximize long-term value in a machine learning system. The system may employ an offline training process and an online training process. In the offline training process, an initial policy is learned to provide a warm start to the online training process. In the online training process, the system applies concurrent reinforcement learning across multiple environments, with the goal of learning efficient policies in real time from in-flight user data in one environment, and applying the learned policies to other environments. With the combination of offline training and online training, the system is able to improve initial performance through the warm start, while adapting to a changing context through concurrent reinforcement learning.