naereen / gym-nes-mario-bros Goto Github PK

Maybe the DQN is failing to learn because it's actually hard to have rules of "what to do when I see that"?

I've been very interested in this project, playing with it locally.
I think it would be helpful to make at least one of the layers (perhaps all of the Conv2D ones) Convolutional 2D LSTM layers. I believe this will help find temporal-spacial relationships within the game screen better and result in better network models.

I really appreciate what you have done there.

But according to what I have seen from your gif, it seems work badly at even this simple game(Mario).

Will it be better to use traditional deep learning tech? for example, supervised learning.

First, we train a model(neural network) by supervised learning(capture screen image as an input, capture keyboard as an output), then we use this model as the base of deep Q network.

I don't know if it's work.

I only knew some basics about keras. But I knew Python very well. I have one month to do this, will you give me some advice when I'm getting stuck later?

Use raw colors from nes (as arbitrary 0-127 values) and not gray from colors

• Right now, having multiple emulators open at the same time can make one bug, when it tries to save things simultaneously than another one.
• It should be easy, use the emulatornumber already used by the pipes, to save to a different csv score file, svg/png model and h5 files.