The　job-shop　scheduling　problem　(JSP)　is　one　of　the　most　famous　production　scheduling　problems,　and　it　is　an　NP-hard　problem.　Reinforcement　learning　(RL),　a　machine　learning　method　capable　of　feedback-based　learning,　holds　great　potential　for　solving　shop　scheduling　problems.　In　this　paper,　the　literature　on　applying　RL　to　solve　JSPs　is　taken　as　the　review　object　and　analyzed　in　terms　of　RL　methods,　the　number　of　agents,　and　the　agent　upgrade　strategy.　We　discuss　three　major　issues　faced　by　RL　methods　for　solving　JSPs:　the　curse　of　dimensionality,　the　generalizability　and　the　training　time.　The　interconnectedness　of　the　three　main　issues　is　revealed　and　the　main　factors　affecting　them　are　identified.　By　discussing　the　current　solutions　to　the　above　issues　as　well　as　other　challenges　that　exist,　suggestions　for　solving　these　problems　are　given,　and　future　research　trends　are　proposed.　©　2024　Elsevier　Ltd