Optimum splitting computing for DNN training through next generation smart networks: a multi-tier deep reinforcement learning approach.