Implement ApeX DQN about tf2rl HOT 6 OPEN

• Change noise level of each explorer

from tf2rl.

• Change noise level of each explorer

edc06fb

from tf2rl.

• Implement flexible target update function
• Reproduce the result of ApeX paper
• Plot evaluation result to TensorBoard

from tf2rl.

Decided to plot learner and evaluator to different two plots at commit dfc0c10 because of followings:

• TensorBoard does not seem to run with more than two files in the same directory
• Avoid sharing tf.contrib.summary because it is often called from both learner and evaluator that might result in significant bottle neck

from tf2rl.

• Unify multi env version and normal version to single script
• Enable to save model periodically

from tf2rl.

Waiting for cpprb implementation of NStepReplayBuffer.

from tf2rl.