Heterogeneous Information Networks (HINs) capture complex relations among entities of various kinds and have been used extensively to improve the effectiveness of various data mining tasks, such as in recommender systems. Many existing HIN-based recommendation algorithms utilize hand-crafted meta-paths to extract semantic information from the networks. These algorithms rely on extensive domain knowledge with which the best set of meta-paths can be selected. For applications where the HINs are highly complex with numerous node and link types, the approach of hand-crafting a meta-path set is too tedious and error-prone. To tackle this problem, we propose the Reinforcement learning-based Meta-path Selection (RMS) framework to select effective meta-paths and to incorporate them into existing meta-path-based recommenders. To identify high-quality meta-paths, RMS trains a reinforcement learning (RL) based policy network(agent), which gets rewards from the performance on the downstream recommendation tasks. We design a HIN-based recommendation model, HRec, that effectively uses the meta-path information. We further integrate HRec with RMS and derive our recommendation solution, RMS-HRec, that automatically utilizes the effective meta-paths. Experiments on real datasets show that our algorithm can significantly improve the performance of recommendation models by capturing important meta-paths automatically.