04 June 2018

Understanding Leela

I ended my previous post Running Leela, with an action:-

Suddenly Black played 34...Qf4, placing the Queen en prise to two pieces. Give that move '??'. I don't understand what happened here and will look into it another time.

First stop: LCZero - Google Groups is an active discussion group where the participants are keen Leela fans with varying knowledge. It didn't take long to find a thread -- What is going on here? (May 29) -- which started with the same question I was asking:-

The above game contains a lot of seemingly absurd moves. Does anybody know why this is happening?

The answer had two parts:-

Training games have randomly selected moves to enable learning. If you want to analyse games played between networks please look for them here: lczero.org/matches.

So training games are the first step in the Leela heirarchy; matches are a second step. Got it. The same post pointed to another introductory resource: leela-chess/wiki/FAQ (github.com).

[Self-play training games] are how Leela Chess Zero trains herself. They are played with extra randomness turned on so it can discover new good (and bad) moves. This means the quality of these games is lower than the match games.

The FAQ should be a good place to start the next time I have a question. Back to the discussion group, how to discover the most relevant threads? I started by looking at recent threads with a high number of responses. Here a few entry level examples:-

  • Delayed gratification, 50-moves rule, discounted rewards, and an exercise (May 21; 73 posts by 14 authors) • 'There are earlier discussions on the LCZero group about Leela Zero aimlessly moving about in a won position, apparently constrained only by the 50-move rule to preserve its anticipation of a win. I don’t like this behaviour, but I have seen it defended by appeal to Zero Knowledge principles; a win is a win and we don’t care for elegance. I like to make a point here, backed by higher authority, that we should care for elegance and that it is appropriate to favour quicker gratification'

  • Super weird match from training games (May 26; 38 posts by 7 authors) • 'lczero.org/game/14801114; Super good start, but from move 29 and anything beyond, this looks like solid randomness.' • 'I wanted to reply that training games are irrelevant, but yeah this doesn't look right. It almost looks like both sides were trying not to win.'

  • Ideas to increase the interest in developing Leela Zero among the general public (May 31; 20 posts by 16 authors) • 'I have been following Leela Zero project for a couple on months now. I am a chess player with little knowledge in AI. I have noticed the number of training games per day is going down, which is going to make the development even slower.'

There are also many technical threads that discuss different configurations for the Leela client or that delve into the structure of the network(s). Let's look at those another time.

No comments: