Network　slicing　technology　can　provide　customized　services　to　many　emerging　applications　with　different　network　requirements,　and　it　is　playing　an　important　role　in　various　fields.　However,　most　of　the　existing　research　work　on　handoff　problems　of　network　slices　use　the　full　buffer　task　model　which　can　not　be　used　to　evaluate　the　performance　of　latency.　Given　latency　is　an　important　metric　for　many　emerging　applications,　the　network　slice　handoff　problem　using　the　finite　buffer　task　model　is　studied　in　this　paper.　Besides,　considering　that　the　increase　of　number　of　users　and　applications　leads　to　the　increasingly　complicated　state　of　the　network,　the　network　slice　handoff　process　is　modeled　as　a　Decentralized　Markov　Decision　Process　(DEC-MDP),　and　a　network　slice　handoff　algorithm　based　on　distributed　deep　reinforcement　learning　is　proposed.　Simulation　results　prove　that　the　algorithm　proposed　in　this　paper　can　significantly　reduce　the　decision　complexity　while　maintaining　a　relatively　superior　overall　system　performance.　©　2023　IEEE.