To　solve　the　problem　of　communication　delay　and　resource　shortage　when　multiple　users　offload　tasks　at　the　same　time　in　mobile　edge　computing　(MEC),　the　deep　reinforcement　learning　algorithm　based　on　non-orthogonal　multiple　access　(NOMA)　technology　was　proposed　to　optimize　users＇　communication　resource　allocation.　Firstly,　the　taboo　tag　deep　Q-network　algorithm　was　used　to　train　the　relationship　between　users　and　subchannels　at　the　users　grouping　stage,　then　the　deep　deterministic　policy　gradient　algorithm　was　used　to　allocate　users　transmission　power　who　sharing　subchannel.　The　simulation　results　display　that　the　proposed　algorithm　perform　more　stable　than　other　reinforcement　learning　and　traditional　algorithm,　moreover,　the　system　sum　rate　have　been　significantly　improved　when　multiple　edge　users　offload　tasks.　©　2022　IEEE.