Deep　Q　Network　(DQN)　takes　the　entire　game　interface　as　input,　makes　use　of　neural　network　to　output　Q　value,　and　maps　it　into　actions.　However,　the　contribution　of　different　game　interfaces　to　Q　value　often　varies,　and　sometimes　only　a　few　interfaces　are　closely　related　to　the　execution　of　agents.　Hence,　we　propose　a　deep　reinforcement　learning　model　based　on　multi-experience　pool　local　state　parallel　Q-Network　(MEPLSPQ-Network),　which　takes　the　advantage　of　multiple　parallel　Q　networks　to　predict　Q　values　collaboratively.　In　this　model,　the　input　of　each　Q　network　is　the　non-overlapping　sub-region　of　the　original　game　interface,　and　subsequently　each　Q　network　will　study　respectively　what　characteristics　different　sub-regions　of　the　game　interface　have.　Experimental　results　indicate　that　the　performance　of　MEPLSPQNetwork　exceeds　that　of　DQN　in　three　various　game　scenes.　©　2019　Association　for　Computing　Machinery.