12 October 2018

Catching Up with Leela

At the end of last week's post, Restoring Adsense, I promised,

After this short series that has little to do with chess, it's time to return to the main subject.

The main subject is chess and the topic to which I want to return is the neural network (NN) engine technology, last seen in The NNs Depart TCEC S13 (August 2018; i.e. Leela and DeusX), and A Leela Surprise in the Nimzo Indian (August 2018; video). Leela is the main player in the NN arena and I once took the time to catalog the online resources that track the progress of this remarkable creation: Finding Leela (June 2018). In the Leela forum we find a recent thread that documents the engine's early days:-

That journey has included three tournaments -- TCEC Seasons 12 & 13 and this year's edition of the CCCC -- of which the last two are TCEC S13 is still underway. Where is the project today? It needs games, lots of games:-

  • 2018-09-26: Worrying drop of contributors • 'There used to be millions of training games per day a few weeks back. Now it is barely 400.000. I remember people in this forum telling me I was stupid to worry about a drop of people who contribute to the training. I fear this is happening now'

Now 400.000 training games per day might sound like enough, but the Leela contributors are beating the bushes to get even more. Here's a recent call for assistance:-

  • 2018-09-26: HOW TO: Contribute games to Leela FOR FREE using Google Cloud and Google Colab • 'I made this post primarily to help people who were wanting to contribute to Leelas development but do not have GPUs, but this is also applicable to those who are already contributing with GPU time. [...] Through Google Cloud you can rent GPUs ranging from K80/P100/V100/etc via your browser with no processing on your end. [...] Google Colab is a similar process whereby you can utilise one of Googles K80 GPUs for free processing again via your browser.

If you're not familiar with GPUs like 'K80 / P100 / V100 / etc.', it's enough to know that they are three evolutions of graphic processors, the number crunchers that all NNs rely on for their numeric intensive computations. To supplement that forum fountain of wisdom, the Leela blog posted its own version:-

I'll certainly be looking into the mechanics of using Google Colab and Google Cloud. Maybe I'll even document my experience here, in my own blog.

No comments: