update tf api for dqn.py

[danielkaifeng]

update dqn.py for workable training. For 12 hours training on server with GPU 1080Ti x2, the loss reduce from 0.008 to 0.003.