TalkChess.com

I don't understand why Leela isn't being trained intensively on simple endgames as part of its training.

Today I've found that the Leela opencl (network=31269, apparently very recent) running on a GPU has serious trouble winning the most basic K+Q vs Q ending. It plays in a completely aimless manner, despite having access to syzygy TBs!

Yes, it can achieve a winning endgame in grandmasterly fashion, but just when things should be super easy they become almost impossibly hard, taking the longest possible time to win such a simple ending. Could be a faulty TB implementation? And why are the depths getting lower and not higher when there are only 3 (three) pieces left on the board?

What do the experts think?

It is an artifact or a side-effect to the current test30 "TB rescoring" implementation. She can convert those endgames quite well, at least better than previous test runs at the same LR, but it takes a little bit of pieces reshuffling and just before the games approaching the 50-move draw she'll start converting those endgames.

carldaman wrote: ↑Sun Nov 18, 2018 3:02 am Today I've found that the Leela opencl (network=31269, apparently very recent) running on a GPU has serious trouble winning the most basic K+Q vs Q ending. It plays in a completely aimless manner, despite having access to syzygy TBs!

But does it look better or worse at endings than networks 1xxxx and 2xxxx?

Hi, as far as I could understand training for 3xxxx NN continues to input whole contributors games (no change here) but substitutes the 5men endgame TB evaluation once such a position is reached for an early accurate and definitive result, not blurred by further randomness (temperature) or lack of knowledge.
Sort of Leela will sooner know for sure the game is won, lost or draw. This will surely help getting skills in early endgames, then aim better targets in late middlegames, and so on.
But I wonder will Leela keep some knowledge to proceed to the expected ending by itself, and convert efficiently? If knowledge gets poor by lack of trained neurons regarding the final steps, will the MCTS do the job?

Ender is a NN trained exclusively on endgames (starting epd’s and 6p TB in self-play). Details here https://github.com/dkappe/leela-chess-w ... ndgame-Net

The thing that’s made the most difference in evaluations has been adding a random half-move clock value of 0-99 to 10% of the starting epd’s. Drawn endgame evaluations have converged toward zero. (Say 0.2 instead of 2.5).