This is probably one of the best videos I've found describing transitioning from tabular Q-learning to using deep Q networks. Other videos seem to sugar coat the topic with terminology that is far too simple to the point where it almost obscures what is actually being discussed.
So, for example we input raw pixels of a frame from the game. Let us say for a 24x24 pixels game, we will have an input layer of 576 neurons, then we will have whatever combination of hidden layers, in the end, we set number of output neurons to be 15 because the size of the action space is 15. Now, how will we able to know which one corresponded to up, down, left, right, etc? You have labelled the output neurons with the corresponding actions but how did you know that?
What kind of silly question is that? Are you learning deep learning for the first time? Read about multiclass classification and maybe come back to this lecture.
One technique is to concatinate 3-4 frames of the game in one input to let the network know about motion. Clearly explained in HuggingFace DQN course notes.
This is probably one of the best videos I've found describing transitioning from tabular Q-learning to using deep Q networks. Other videos seem to sugar coat the topic with terminology that is far too simple to the point where it almost obscures what is actually being discussed.
Great staff! Thank you very much.
So, for example we input raw pixels of a frame from the game. Let us say for a 24x24 pixels game, we will have an input layer of 576 neurons, then we will have whatever combination of hidden layers, in the end, we set number of output neurons to be 15 because the size of the action space is 15. Now, how will we able to know which one corresponded to up, down, left, right, etc? You have labelled the output neurons with the corresponding actions but how did you know that?
What kind of silly question is that? Are you learning deep learning for the first time? Read about multiclass classification and maybe come back to this lecture.
Pl see sentdex video , it is more clear with code
Where are we passing actions as input ? Passing only states will not help to calculate q values . I think some important part is missing in video.
One technique is to concatinate 3-4 frames of the game in one input to let the network know about motion. Clearly explained in HuggingFace DQN course notes.