model.add(Flatten(input_shape=(1,) + env.observation_space.shape)) about keras-rl

[matthiasplappert]

Exactly, it is related to the window length. Let's assume that your observations have shape (10, 20). If window_length=4, the input to the network would be of shape (4, 10, 20). If you use flatten or not depends on what you want to achieve. In the Atari example, the convolution is performed before flattening since the same set of filter should be used for each frame. In other examples, it might make more sense to flatten immediately.

In the case of the cartpole example, the window length is 1 (since the observation already contains velocities etc.). Flattening the input thus simply reduces the dimension by one and otherwise keeps the observation identical.

Does this clarify things?

[matthiasplappert]

Closing this since the question has (hopefully) been answered. If you this needs further clarification, please re-open the issue.

[yongduek]

Hello, the question was answered; sorry for not responding due to my tight
work schedule. I should have left some message.
Thanks a lot for the answer.

Yongduek Seo
Professor
[phone]
Department of Media Technology
Sogang University, Korea

On 6 September 2016 at 22:21, Matthias Plappert [email protected]
wrote:

Closed #19 https://github.com/matthiasplappert/keras-rl/issues/19.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
https://github.com/matthiasplappert/keras-rl/issues/19#event-779320444,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AD3huKH15Fc3-hW7tWFlwbmrC39T88Hnks5qnWjygaJpZM4JtAFK
.