Network slicing enables an infrastructure provider (InP) to support heterogeneous 5G services over a common platform (i.e., by creating a customized slice for each service). Once in operation, slices can be dynamically scaled up/down to match the variation of their service requirements. An InP generates revenue by accepting a slice request. If a slice cannot be scaled up when required, an InP has to also pay a penalty (proportional to the level of service degradation). It becomes then crucial for an InP to decide which slice requests should be accepted/rejected in order to increase its net profit. This paper presents a slice admission strategy based on reinforcement learning (RL) in the presence of services with different priorities. The use case considered is a 5G flexible radio access network (RAN), where slices of different mobile service providers are virtualized over the same RAN infrastructure. The proposed policy learns which are the services with the potential to bring high profit (i.e., high revenue with low degradation penalty), and hence should be accepted. The performance of the RL-based admission policy is compared against two deterministic heuristics. Results show that in the considered scenario, the proposed strategy outperforms the benchmark heuristics by at least 23%. Moreover, this paper shows how the policy is able to adapt to different conditions in terms of 1) slice degradation penalty versus slice revenue factors, and 2) proportion of high versus low priority services.