Add distributed environment for training about dream-go HOT 3 CLOSED

It would be nice if could also see which weights were used to generate each self-play game and feature. So I might want to add the generation field (can be null) to the database and extend the REST protocol to allow for some way to include the generation in there (maybe as a header?), we also need some way to retrieve the latest generation using the REST-ful API.

from dream-go.

How many features do we need to include during training. More is better, but being realistic, where does the diminishing returns start kicking in?

• DeepMind includes the 500,000 most recent games (but extract multiple features from each game?)
• Thinking Fast and Slow with Deep Learning and Tree Search use DAGGER which trains on the entire dataset.

The network that DeepMind were training is roughly twice as big as ours is, so we can probably get away with half of their number of games. But we could also follow the DAGGER approach and train on everything.

I have some concern with training on everything since some early games can be very bad and imitating them will not get us anything good. So we will probably go with the DeepMind approach and train on the 250,000 most recent games. This is a nice number also because that is about the number of games we can produce in a day using two GPU's.

from dream-go.

I checked in the mentioned images, see the the docker directory for all the details.

from dream-go.