Help to replicate the results

[vishalkumarchaudhary]

Hello,
Since the repository does not contain enough information fo how to start. I wanted to replicate the result of the above game. Can you tell us the detail of how to start training?

Thanks in advance

[richemslie]

Hi,
Sorry for the delay. Which "above game" were you referring to? :)

I'd recommend breakthrough 6x6 as it a very fast game to train from scratch on a single GPU. Also, if it is one of the games I have already trained, I can upload a recent model / give access to the training data.

Once you've decided on which game to train/test, the basic steps to familiarise with the codebase(s) would be something like:

1. run a self-play mini tournament testing the configuration options to the PUCT player (thereby testing installation, dependencies, tensorflow/GPU, etc).
2. Use the existing training data (or even random training data if no existing data exists) to train a fresh model, supervised. This will get you familiar with the various options, network model choices, post-processing database.
3. Once familiar with those two, then the final step is to set up a config for the client/server to automate self-play and training. This will be much easier if you have completed the first 2 steps.

Obviously, each step will raise a lot more questions if you go ahead. Hopefully, during the process, you can aid with writing documentation so it would be easier for others to get going.

Finally, a word of warning - some configuration options will seem a bit nebulous given the number of users of this project == 1. Some of these options need to be moved or renamed. Having someone else other than me train a game, will force me to explain these options and thereby rename them to be clearer. So I guess that will be part of the process! :)