Use the training end_e as the evaluation(..., epsilon=end_e) for atari

Use the training end_e as the evaluation(..., epsilon=end_e) for atari #879