As　a　promising　technique,　mobile　edge　computing　(MEC)　has　attracted　significant　attention　from　both　academia　and　industry.　However,　the　offloading　decision　for　computing　tasks　in　MEC　is　usually　complicated　and　intractable.　In　this　paper,　we　propose　a　novel　framework　for　offloading　decision　in　MEC　based　on　Deep　Reinforcement　Learning　(DRL).　We　consider　a　typical　network　architecture　with　one　MEC　server　and　one　mobile　user,　in　which　the　tasks　of　the　device　arrive　as　a　flow　in　time.　We　model　the　offloading　decision　process　of　the　task　flow　as　a　Markov　Decision　Process　(MDP).　The　optimization　object　is　minimizing　the　weighted　sum　of　offloading　latency　and　power　consumption,　which　is　decomposed　into　the　reward　of　each　time　slot.　The　elements　of　DRL　such　as　policy,　reward　and　value　are　defined　according　to　the　proposed　optimization　problem.　Simulation　results　reveal　that　the　proposed　method　could　significantly　reduce　the　energy　consumption　and　latency　compared　to　the　existing　schemes.