Effect of Kostrikov's third convolutional layer with 32 instead of 64 channels

assigned to @dl337788

changed milestone to %Official dates

added Point of Interest label

Kostrikov also has a linear layer with input_size=512, output_size=action_space.n which is frozen after orthogonal initialization. It is not part of the model itself, which is fed to the Adam optimizer, but of a wrapper around the categorical distribution, which is not optimized at all.

added In Review label

closed

removed In Review label