This is a first step towards allowing serialization/deserialization to arrays and strings. This will allow us to send policy weights across the network without sending them to disk first. It will also give us an easier interface to use when we want to modify policy weights using evolutionary update steps.
I didn't change any of the logic around replacing global/ names with local/, but when i tested this pattern, I realize the existing code already has a problem. If there are any global/ variable name when restoring saved state, the code will attempt to load those global variables, and they won't exist in the same state. I am not sure how this currently working in master. This code doesn't modify any of that logic however.
This is a first step towards allowing serialization/deserialization to arrays and strings. This will allow us to send policy weights across the network without sending them to disk first. It will also give us an easier interface to use when we want to modify policy weights using evolutionary update steps.
I didn't change any of the logic around replacing
global/
names withlocal/
, but when i tested this pattern, I realize the existing code already has a problem. If there are anyglobal/
variable name when restoring saved state, the code will attempt to load those global variables, and they won't exist in the same state. I am not sure how this currently working in master. This code doesn't modify any of that logic however.