Training the trainer: how is it done for Stockfish?

Discussion of chess software programming and technical issues.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
syzygy
Posts: 4455
Joined: Tue Feb 28, 2012 10:56 pm

Re: Training the trainer: how is it done for Stockfish?

Post by syzygy » Fri Mar 15, 2019 10:08 pm

Robert Pope wrote:
Fri Mar 15, 2019 6:11 pm
When these positions arise in the middle game, Leela will often miss the correct continuation, just like it does in the endgame. The question then becomes "are positions that require extremely precise play equally likely to arise in the middle game as they are in the endgame?" I believe the answer to that is "no". I think that early in a chess game, there tend to be many workable strategies that can be followed, so positions requiring a single continuation of perfect play are not common. As you get closer to the endgame, more situations arise where the wrong move can make the difference between a draw and a loss. e.g. if you are one move too late moving your king toward an enemy pawn, you will lose a pawn race, and be down a queen.
And I guess the number of workable strategies when playing a classical engine is also a function of how "perfectly" the classical engine can play the position. In the endgame, a classical engine can often play perfectly, punishing any subtle error its opponent makes. In the middle game, a classical engine cannot fully "solve" the position and is likely to make subtle mistakes itself.

Where NNs could get into trouble are middle game positions with a relatively deep tactic that a classical engine can spot but an NN-based engine not. LC0's opening play is apparently good enough to usually avoid positions with such tactical traps.

JollyJoker
Posts: 3
Joined: Sat Feb 16, 2019 7:43 pm
Full name: Lennart Qvarnström

Re: Training the trainer: how is it done for Stockfish?

Post by JollyJoker » Sat Mar 16, 2019 9:52 am

Comparing n-man to n+1 man, does n+1 man have a higher share of moves that don't worsen the outcome? Assuming we're considering that as the definition of a mistake, rather than having an opponent that can actually exploit the mistake.

syzygy
Posts: 4455
Joined: Tue Feb 28, 2012 10:56 pm

Re: Training the trainer: how is it done for Stockfish?

Post by syzygy » Sat Mar 16, 2019 11:38 pm

JollyJoker wrote:
Sat Mar 16, 2019 9:52 am
Comparing n-man to n+1 man, does n+1 man have a higher share of moves that don't worsen the outcome? Assuming we're considering that as the definition of a mistake, rather than having an opponent that can actually exploit the mistake.
I would say: compared with positions with n men, positions with n+1 men have a higher share of moves that aren't reliably evaluated by a classical engine (given the same time per move).

NNs are good at fuzzy things. With more pieces left on the board, chess is more fuzzy for classical engines, so NNs have better chances to outplay classical engines.

Post Reply