What is aim of Frame-skipping technology?

[qq406...@gmail.com]

In the Section 5 of the paper  "Playing Atari With Deep Reinforcement Learning", the author use a frame-skipping technology.

I don't know why we select action every 4 frames instead of 5 or 8 frames.

My thought is that we need the time of 4 frames to training the CNN and we can implement real-time playing game.

However,I am not sure the time of 4 frames is enough to achieve the training.If one frame is 1/12 second ,the 4 frames is about 0.33 second.

The time in training network seem not just consume just 0.33 second.

If the number of frame-skipping is not connected with real-time playing game,

my question is how the  Q network implement combination the training and selection.

Does anyone can explain this question?

[Maciej Jaśkowski]

Hi there,

Frame skiping  is used primarily to speed up  training.

4 consecutive frames are very similar to each other and arguably don't provide new information for training.

You could skip 6 or 8 frames with similar results but according to the report in nature paper 4 seemed to work best across range of games.

Good luck!
Maciej

--
You received this message because you are subscribed to the Google Groups "Deep Q-Learning" group.
To unsubscribe from this group and stop receiving emails from it, send an email to deep-q-learni...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/deep-q-learning/399e7e6e-37a6-4249-94ab-002d5599cf48%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.