A key function of a quantum internet is the generation of entangled links between devices. The quality of these links decays over time due to interaction with the outside world. Various protocols exist for generating these links. The main trade-off to be considered when choosing a protocol is between the chance of producing the link and the quality of the generated link. Selecting the optimal protocol is a complex and computationally intensive task, especially as the scale of the problem increases. While analytical solutions are feasible for small-scale systems, they become impractical for larger networks due to their computational demands. In addition, the rate at which the generated links decay is usually assumed to be static, when in reality it typically fluctuates over time in a process called parameter drift. In this paper, we evaluate the effectiveness of reinforcement learning approaches to optimizing protocol selection for different models of this problem. Models are provided for a static rate of decay, for a drifting rate of decay, and for the phenomenon of induced decoherence. Different reinforcement learning agents are trained on all of these models, and the results are compared. The criterion for effectiveness is the speed at which the requested number of entangled links can be generated. A negligible effect on performance is detected, showing that models are able to adapt to the different sources of instability studied. We also provide methods to model this problem and identify promising directions for future research.