but I am new to Python and don't have all the required libraries on hand. To gear up to run this might be a bit of work, which is OK but it would be nice to know if this is the right starting point or if other code is preferable.
If I get a bot running I would be happy to offer it for others to download.
Or would any of the people running bots here be kind enough to offer downloads (hint, hint).
I do not have any spare time but if by any chance I will have, then I would like to help. I would like to learn some machine learning and generally Im a python backend developer so maybe Im able to at least understand some of the existing code ;)
Just wanted to point out that gzero has been available to "download" since the first line of code was written. It is a GGP bot, not a hex bot however, and it isn't exactly double click and away you go.
I was thinking if 15x15 comes to fruition here on LG, perhaps anyone here would like to be part of a collaborative effort for a new bot? I am thinking 95% pure python, minimal dependencies and keeping things as simple as possible.
So I managed to run the script that TomAce has linked. It has some ugly hard-coded constants which I corrected, but it doesn't have a requirements file so I'm not sure if I use proper libraries' versions. Generally it runs, but the problem is that it takes up all my memory and dies.
Does anybody have an experience in how to manage tensorflow 2.0 memory usage?
I found that the script could be optimized to go significantly faster with just the code optimization. Also the huge RAM consumption is avoidable. I would maybe find time to do all that, but I hate to be stuck at the configuration level and I failed a couple attempts to have tensorflow use my GPU.
If there are people who have configured GPU and tensorflow in the past I will appreciate all advice. I'm using nvidia gtx 850m and linux mint.
As it stands the gzero code (galvanise_zero) is extremely fiddly, and is based on the now very inactive Stanford General Game Playing project and ggp-base java library. Not to mention it is using python 2, and a very old version of TensorFlow.
So I decided to start afresh with a new project, with the goal of training hex 11x11 using Deepmind's MuZero method. I'll use most of what I learned with gzero, so other game types will be easy to add.
However, not sure I will have much time over the holidays to work on it, but can follow here if you like