Free engines with positional understanting at depth 10.

Discussion of anything and everything relating to chess playing software and machines.

Moderator: Ras

syzygy
Posts: 5898
Joined: Tue Feb 28, 2012 11:56 pm

Re: Free engines with positional understanting at depth 10.

Post by syzygy »

FireDragon761138 wrote: Tue Feb 03, 2026 8:35 am
syzygy wrote: Tue Feb 03, 2026 7:23 am
FireDragon761138 wrote: Tue Feb 03, 2026 3:30 am
syzygy wrote: Tue Feb 03, 2026 1:16 am
FireDragon761138 wrote: Mon Feb 02, 2026 10:31 pm Preliminary testing seems to suggest SF 16.1 might be somewhat more stable in terms of evaluations than Stockfish 17.1, which generally means it's going to be better at demonstrating stable plans or ideas. SF17.1 seems to have the LMR algorithm being more directly influenced by the NNUE output, which inclines it towards more aggressive search for sharp tactics.
Any PV output by an engine is as good as any other (while understanding that PVs for higher depths will generally have more accurate moves).
If the engine switches a lot you just get more PVs, which HELPS the user to understand the position.
To always get multiple PVs, use multipv=N. This option is specifically intended for analysis.
The more PVs the better.

The idea that "stable evaluations" and "stable PVs" are inherently a good thing is fundamentally flawed. Many positions simply inherently have many different possible continuations. You as the player can control your own moves but not the moves of the opponent. And even when you can choose a move, you want to know whether there are multiple promising moves.

Of course to really understand a position the user should also study the various possible lines separately.
A stable positional evaluation is better in the sense of gesturing towards actual positional ideas or themes. It's the difference between a positionally sound move, and one that's so hyper-contextual, that there's nothing to learn from it. They are not the same, and one facilitates much more generalisable learning than the other.
No, it is not. Those statements are based on nothing and essentially meaningless. An evaluation does not "gesture at positional ideas". Studying a PV might uncover an idea, but again, any PV is as good or bad as any other for this. PVs that vary over search iterations simply show more possible variations.
Your just dismissing my point with Positivist-style hand-waving "there's no meaning here, only data... and the more, the better". Yeah, in which case, why use engines at all for analysis? If relevance and context aren't important in deciding what engine to use for a specific task, why not just stick with Sargon? It's about as useful for the average human player if all you want is "show me the best tactical move that I can learn something from"
I gave plenty of explanation why a stable PV is irrelevant for analysis purposes, and you simply ignore that explanation. Why is a PV at depth 30 less useful for analysis if it differs from the PV at depth 29? Can you answer that?

Amd what do you mean by an evaluation that "gestures"? An evaluation is just a number. Numbers don't gesture. Numbers are not related to positional ideas. If you are unable to express your ideas in concrete language, then maybe the ideas aren't very helpful.
FireDragon761138
Posts: 80
Joined: Sun Dec 28, 2025 7:25 am
Full name: Aaron Munn

Re: Free engines with positional understanting at depth 10.

Post by FireDragon761138 »

syzygy wrote: Tue Feb 03, 2026 12:03 pm
FireDragon761138 wrote: Tue Feb 03, 2026 8:35 am
syzygy wrote: Tue Feb 03, 2026 7:23 am
FireDragon761138 wrote: Tue Feb 03, 2026 3:30 am
syzygy wrote: Tue Feb 03, 2026 1:16 am
FireDragon761138 wrote: Mon Feb 02, 2026 10:31 pm Preliminary testing seems to suggest SF 16.1 might be somewhat more stable in terms of evaluations than Stockfish 17.1, which generally means it's going to be better at demonstrating stable plans or ideas. SF17.1 seems to have the LMR algorithm being more directly influenced by the NNUE output, which inclines it towards more aggressive search for sharp tactics.
Any PV output by an engine is as good as any other (while understanding that PVs for higher depths will generally have more accurate moves).
If the engine switches a lot you just get more PVs, which HELPS the user to understand the position.
To always get multiple PVs, use multipv=N. This option is specifically intended for analysis.
The more PVs the better.

The idea that "stable evaluations" and "stable PVs" are inherently a good thing is fundamentally flawed. Many positions simply inherently have many different possible continuations. You as the player can control your own moves but not the moves of the opponent. And even when you can choose a move, you want to know whether there are multiple promising moves.

Of course to really understand a position the user should also study the various possible lines separately.
A stable positional evaluation is better in the sense of gesturing towards actual positional ideas or themes. It's the difference between a positionally sound move, and one that's so hyper-contextual, that there's nothing to learn from it. They are not the same, and one facilitates much more generalisable learning than the other.
No, it is not. Those statements are based on nothing and essentially meaningless. An evaluation does not "gesture at positional ideas". Studying a PV might uncover an idea, but again, any PV is as good or bad as any other for this. PVs that vary over search iterations simply show more possible variations.
Your just dismissing my point with Positivist-style hand-waving "there's no meaning here, only data... and the more, the better". Yeah, in which case, why use engines at all for analysis? If relevance and context aren't important in deciding what engine to use for a specific task, why not just stick with Sargon? It's about as useful for the average human player if all you want is "show me the best tactical move that I can learn something from"
I gave plenty of explanation why a stable PV is irrelevant for analysis purposes, and you simply ignore that explanation. Why is a PV at depth 30 less useful for analysis if it differs from the PV at depth 29? Can you answer that?

Amd what do you mean by an evaluation that "gestures"? An evaluation is just a number. Numbers don't gesture. Numbers are not related to positional ideas. If you are unable to express your ideas in concrete language, then maybe the ideas aren't very helpful.
A PV is part of the engine's analysis or evaluation of a position as output. Perhaps you are thinking from a strictly programming/coding perspective, but I was thinking in terms of how engines are actually used.
syzygy
Posts: 5898
Joined: Tue Feb 28, 2012 11:56 pm

Re: Free engines with positional understanting at depth 10.

Post by syzygy »

FireDragon761138 wrote: Tue Feb 03, 2026 1:04 pm
syzygy wrote: Tue Feb 03, 2026 12:03 pm
FireDragon761138 wrote: Tue Feb 03, 2026 8:35 am
syzygy wrote: Tue Feb 03, 2026 7:23 am
FireDragon761138 wrote: Tue Feb 03, 2026 3:30 am
syzygy wrote: Tue Feb 03, 2026 1:16 am
FireDragon761138 wrote: Mon Feb 02, 2026 10:31 pm Preliminary testing seems to suggest SF 16.1 might be somewhat more stable in terms of evaluations than Stockfish 17.1, which generally means it's going to be better at demonstrating stable plans or ideas. SF17.1 seems to have the LMR algorithm being more directly influenced by the NNUE output, which inclines it towards more aggressive search for sharp tactics.
Any PV output by an engine is as good as any other (while understanding that PVs for higher depths will generally have more accurate moves).
If the engine switches a lot you just get more PVs, which HELPS the user to understand the position.
To always get multiple PVs, use multipv=N. This option is specifically intended for analysis.
The more PVs the better.

The idea that "stable evaluations" and "stable PVs" are inherently a good thing is fundamentally flawed. Many positions simply inherently have many different possible continuations. You as the player can control your own moves but not the moves of the opponent. And even when you can choose a move, you want to know whether there are multiple promising moves.

Of course to really understand a position the user should also study the various possible lines separately.
A stable positional evaluation is better in the sense of gesturing towards actual positional ideas or themes. It's the difference between a positionally sound move, and one that's so hyper-contextual, that there's nothing to learn from it. They are not the same, and one facilitates much more generalisable learning than the other.
No, it is not. Those statements are based on nothing and essentially meaningless. An evaluation does not "gesture at positional ideas". Studying a PV might uncover an idea, but again, any PV is as good or bad as any other for this. PVs that vary over search iterations simply show more possible variations.
Your just dismissing my point with Positivist-style hand-waving "there's no meaning here, only data... and the more, the better". Yeah, in which case, why use engines at all for analysis? If relevance and context aren't important in deciding what engine to use for a specific task, why not just stick with Sargon? It's about as useful for the average human player if all you want is "show me the best tactical move that I can learn something from"
I gave plenty of explanation why a stable PV is irrelevant for analysis purposes, and you simply ignore that explanation. Why is a PV at depth 30 less useful for analysis if it differs from the PV at depth 29? Can you answer that?

Amd what do you mean by an evaluation that "gestures"? An evaluation is just a number. Numbers don't gesture. Numbers are not related to positional ideas. If you are unable to express your ideas in concrete language, then maybe the ideas aren't very helpful.
A PV is part of the engine's analysis or evaluation of a position as output. Perhaps you are thinking from a strictly programming/coding perspective, but I was thinking in terms of how engines are actually used.
I asked clear questions. I understand you are unable to answer them. I did not expect anything else.
Ciekce
Posts: 206
Joined: Sun Oct 30, 2022 5:26 pm
Full name: Conor Anstey

Re: Free engines with positional understanting at depth 10.

Post by Ciekce »

FireDragon761138 wrote: Mon Feb 02, 2026 10:31 pm Preliminary testing seems to suggest SF 16.1 might be somewhat more stable in terms of evaluations than Stockfish 17.1, which generally means it's going to be better at demonstrating stable plans or ideas. SF17.1 seems to have the LMR algorithm being more directly influenced by the NNUE output, which inclines it towards more aggressive search for sharp tactics.
Why are you so laser-focused on being a walking case study on the Dunning-Kruger effect? None of this is true.
FireDragon761138
Posts: 80
Joined: Sun Dec 28, 2025 7:25 am
Full name: Aaron Munn

Re: Free engines with positional understanting at depth 10.

Post by FireDragon761138 »

Ciekce wrote: Fri Feb 06, 2026 11:34 pm
FireDragon761138 wrote: Mon Feb 02, 2026 10:31 pm Preliminary testing seems to suggest SF 16.1 might be somewhat more stable in terms of evaluations than Stockfish 17.1, which generally means it's going to be better at demonstrating stable plans or ideas. SF17.1 seems to have the LMR algorithm being more directly influenced by the NNUE output, which inclines it towards more aggressive search for sharp tactics.
Why are you so laser-focused on being a walking case study on the Dunning-Kruger effect? None of this is true.
You keep asserting that I'm an idiot without any evidence. Have you ever heard the saying that "every accusation is a confession"?

Aside from what I've already discussed about relatively stable evaluations from ply to ply gesturing at conceptual self-similarity and therefore stability, on the practical side, if engine stability doesn't matter for understanding strategic ideas in the PV of an engine, why does the Dragon engine have a dynamism setting that allows a user to tune the engine either for elo strength or human comprehension?
lucario6607
Posts: 44
Joined: Sun May 19, 2024 5:44 am
Full name: Kolby Mcgowan

Re: Free engines with positional understanting at depth 10.

Post by lucario6607 »

FireDragon761138 wrote: Tue Feb 03, 2026 1:04 pm
syzygy wrote: Tue Feb 03, 2026 12:03 pm
FireDragon761138 wrote: Tue Feb 03, 2026 8:35 am
syzygy wrote: Tue Feb 03, 2026 7:23 am
FireDragon761138 wrote: Tue Feb 03, 2026 3:30 am
syzygy wrote: Tue Feb 03, 2026 1:16 am
FireDragon761138 wrote: Mon Feb 02, 2026 10:31 pm Preliminary testing seems to suggest SF 16.1 might be somewhat more stable in terms of evaluations than Stockfish 17.1, which generally means it's going to be better at demonstrating stable plans or ideas. SF17.1 seems to have the LMR algorithm being more directly influenced by the NNUE output, which inclines it towards more aggressive search for sharp tactics.
Any PV output by an engine is as good as any other (while understanding that PVs for higher depths will generally have more accurate moves).
If the engine switches a lot you just get more PVs, which HELPS the user to understand the position.
To always get multiple PVs, use multipv=N. This option is specifically intended for analysis.
The more PVs the better.

The idea that "stable evaluations" and "stable PVs" are inherently a good thing is fundamentally flawed. Many positions simply inherently have many different possible continuations. You as the player can control your own moves but not the moves of the opponent. And even when you can choose a move, you want to know whether there are multiple promising moves.

Of course to really understand a position the user should also study the various possible lines separately.
A stable positional evaluation is better in the sense of gesturing towards actual positional ideas or themes. It's the difference between a positionally sound move, and one that's so hyper-contextual, that there's nothing to learn from it. They are not the same, and one facilitates much more generalisable learning than the other.
No, it is not. Those statements are based on nothing and essentially meaningless. An evaluation does not "gesture at positional ideas". Studying a PV might uncover an idea, but again, any PV is as good or bad as any other for this. PVs that vary over search iterations simply show more possible variations.
Your just dismissing my point with Positivist-style hand-waving "there's no meaning here, only data... and the more, the better". Yeah, in which case, why use engines at all for analysis? If relevance and context aren't important in deciding what engine to use for a specific task, why not just stick with Sargon? It's about as useful for the average human player if all you want is "show me the best tactical move that I can learn something from"
I gave plenty of explanation why a stable PV is irrelevant for analysis purposes, and you simply ignore that explanation. Why is a PV at depth 30 less useful for analysis if it differs from the PV at depth 29? Can you answer that?

Amd what do you mean by an evaluation that "gestures"? An evaluation is just a number. Numbers don't gesture. Numbers are not related to positional ideas. If you are unable to express your ideas in concrete language, then maybe the ideas aren't very helpful.
A PV is part of the engine's analysis or evaluation of a position as output. Perhaps you are thinking from a strictly programming/coding perspective, but I was thinking in terms of how engines are actually used.
Engines like stockfish are not made to teach humans so clearly your thinking is wrong