Closed Joshua-Chin closed 7 years ago
The tf.Variable is created in the linear
function here: https://github.com/openai/universe-starter-agent/blob/master/model.py#L37. It is modified by the trainer to maximize reward. It is normal for the initializer of a variable to be a constant.
normalized_columns_initializer
returns atf.constant
instead of atf.Variable
. This means that the weights of the linear layers will be held constant through training. I'm not sure if this was intentional.