txizzle / drl Goto Github PK

I noticed that you have the lines:

X_train = X_train.reshape(X_train.shape[0], obs_data.shape[1], 1)
...

model.add(Dense(128, activation='relu', input_shape=(obs_data.shape[1], 1)))
model.add(Flatten())

Why do you make the array 3D only to almost immediately flatten it? I'm not really seeing the motivation.

Hi,

when running the DAgger algorithm, do you retrain the network with the current weights loaded or do you retrain "from scratch"?

The DAgger paper is not really clear about the practical implementation.

Thanks.