Blunder v8.0.0 underestimates danger ? about blunder

[deanmchris]

Hi, @tissatussa thanks for the report.

Running the position myself, Blunder 8.0.0, with a fresh start, finds g6 instantly, but erroneously switches to b6 at depth 18. I think this is likely due to a new pruning technique I added in Blunder known as razoring, which you may be familiar with. Removing the razoring code and researching the position, Blunder finds g6 instantly and sticks with it. So Blunder blundering here was likely caused by some faulty razoring, which made it blind to the force of Qh5.

Trying to tune your search parameters well is a million-dollar question, and in hyper-bullet testing over about 4500 games, razoring proved to be a net win, but in positions like the one you showed, it's likely too aggressive I'll need to work on tuning it down for the next release of Blunder.

Unfortunately pruning always comes with risks, even if they're a net win. The eventual goal of course is to reach an acceptable balance between blindness in some scenarios, but overall search strength.

Now hopefully this one game isn't indictive of how well Blunder 8.0.0. plays in general, but if it is, I'll have some work to...to say the least.

[mkd]

Hi, and sorry for getting in the middle of the conversation, but: I just noticed that, in your moves loop you have a very aggressive condition where any move with SEE<0 will be immediately pruned. Here's the link to the code in search.go: https://github.com/algerbrex/blunder/blob/527906ccccdb0c603dc62e9ea1a885c0c7c333f3/engine/search.go#L637 I tried that pruning condition on my own engine and it starts making huge blunders when there's a potential threat or mate, and it actually depends on a negative capture. If you try removing that condition, the engine (also Blunder 8.0) starts playing so much better. Regards, --Claudio

…

-- Claudio M. Camacho Strategic Marketing Management // Business Strategy & Leadership // Executive Board Advisor tel: [phone] web: www.claudiocamacho.io twitter: twitter.com/claudiomkd linkedin: http://www.linkedin.com/in/claudiocamacho medium: ***@***.***

On Thu, Jun 30, 2022 at 8:08 PM Christian Dean ***@***.***> wrote: Hi, @tissatussa <https://github.com/tissatussa> thanks for the report. Running the position myself, Blunder 8.0.0, with a fresh start, finds g6 instantly, but erroneously switches to b6 at depth 18. I think this is likely due to a new pruning technique I added in Blunder known as razoring, which you may be familiar with. Removing the razoring code and researching the position, Blunder finds g6 instantly and sticks with it. So Blunder blundering here was likely caused by some faulty razoring, which made it blind to the force of Qh5. Trying to tune your search parameters well is a million-dollar question, and in hyper-bullet testing over about 4500 games, razoring proved to be a net win, but in positions like the one you showed, it's likely too aggressive I'll need to work on tuning it down for the next release of Blunder. — Reply to this email directly, view it on GitHub <#20 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAEMXDCXLD6FUEPZIELOPPTVRXICHANCNFSM52IW7DEA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[deanmchris]

Hi @mkd thanks for the heads up.

You're correct. In quiesecene search, I am pruning all moves that have an SEE score of 0. As I mentioned to tissatusa, the hope is that the blindness that comes with doing such an approach is minimal, and mostly out weighed by the strength gained by being able to search good lines deeper on average. And if I remember correctly from my hyperbullet testing, SEE pruning proved to be an overall win.

You mentioned Blunder started playing much better after SEE was removed. Did you mean you also did some testing to determine the Elo gain/loss from removing SEE pruning from Blunder? Could you share those results? Thanks!

[mkd]

Hi Christian, No, I haven't done any proper testing. I just happened to implement my own SEE and got a lot of ideas from Blunder's code, because you have a nice and clean implementation (in my opinion). Then, I checked where do you use SEE() for pruning and, within qsearch it's completely fine. However, when I tried to prune also in the general search, at the beginning of the "moves loop", my engine started to miss out on tactical positions where there were deep threats. I read the message from Roelof and tried the same position: r1bq2k1/pp3rpp/2p2b2/3p4/1B6/1R1B3P/P1P2PP1/3QR1K1 b - - 0 20 After that, I realized having SEE() as a pruning decision in the general search is too harsh and will miss out on positions where there are long-term deep threats that are not imminent. I didn't do anything else than that. But feel free to try removing SEE() from your moves loop in the general search, and see what happens. I also tested my own engine and the amount of nodes saved, in my case, is almost neglectible. So having SEE() pruning or not in the general search doesn't help my engine at all. Therefore, I only kept it in the qsearch. Hope this explanation helps you a bit more. By the way, thank you for writing Blunder, I've learned a lot from your code and introduced a few ideas into my own engine, adapting them to C++ code. Keep up the great work!

…

-- Claudio M. Camacho Strategic Marketing Management // Business Strategy & Leadership // Executive Board Advisor tel: [phone] web: www.claudiocamacho.io twitter: twitter.com/claudiomkd linkedin: http://www.linkedin.com/in/claudiocamacho medium: ***@***.***

On Fri, Jul 1, 2022 at 12:48 AM Christian Dean ***@***.***> wrote: Hi @mkd <https://github.com/mkd> thanks for the heads up. You're correct. In *quiesecene* search, I am pruning all moves that have an SEE score of 0. As I mentioned to tissatusa, the hope is that the blindness that comes with doing such an approach is minimal, and mostly out weighed by the strength gained by being able to search good lines deeper on average. And if I remember correctly from my hyperbullet testing, SEE pruning proved to be an overall win. You mentioned Blunder started playing much better after SEE was removed. Did you mean you also did some testing to determine the Elo gain/loss from removing SEE pruning from Blunder? Could you share those results? Thanks! — Reply to this email directly, view it on GitHub <#20 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAEMXDAKQ6RGFANNWV73NYDVRYI2LANCNFSM52IW7DEA> . You are receiving this because you were mentioned.Message ID: ***@***.***>