I ran a couple of Patricia - Winter games. The games have gotten far more competitive! Last time I had tested, I had to put fairly large time handicaps for games to get interesting, and it felt like Winter was usually gradually outplaying, but would occasionally blunder tactically. Now it feels like there are times where Patricia better evaluates some kind of dynamic compensation.
Patricia is still weaker than Winter, but it is close enough that it doesn't feel necessary to have a time handicap. In the test games, I ran Winter with high levels of contempt, hopefully bringing the engines fairly close in strength. Unsurprisingly, I have not seen a draw in my small sample size yet. The following are the two games played at 95 and 99 contempt respectively with Patricia beating 99 contempt and losing a tense fight against 95 contempt.
I'm not sure about this statement though as it got into a lot of trouble against a 2300 Elo engine: "It's worth noting that Patricia is still well into superhuman territory, with an estimated CCRL elo of 3100-3150; she'll lose badly to top engines, but will still crush any human."
Werewolf wrote: ↑Wed Apr 03, 2024 11:49 pm
What an incredibly fun engine!
I'm not sure about this statement though as it got into a lot of trouble against a 2300 Elo engine: "It's worth noting that Patricia is still well into superhuman territory, with an estimated CCRL elo of 3100-3150; she'll lose badly to top engines, but will still crush any human."
Thanks for the kind words!
She's absolutely superhuman, I can guarantee you that, however I wouldn't be surprised if she sometimes had a propensity to mess up with some ridiculous sacrifice against a weaker engine.
Werewolf wrote: ↑Wed Apr 03, 2024 11:49 pm
What an incredibly fun engine!
I'm not sure about this statement though as it got into a lot of trouble against a 2300 Elo engine: "It's worth noting that Patricia is still well into superhuman territory, with an estimated CCRL elo of 3100-3150; she'll lose badly to top engines, but will still crush any human."
Thanks for the kind words!
She's absolutely superhuman, I can guarantee you that, however I wouldn't be surprised if she sometimes had a propensity to mess up with some ridiculous sacrifice against a weaker engine.
Do you think you’ll make it multi threaded in the future? (I can only set to 1 thread)
Werewolf wrote: ↑Wed Apr 03, 2024 11:49 pm
What an incredibly fun engine!
I'm not sure about this statement though as it got into a lot of trouble against a 2300 Elo engine: "It's worth noting that Patricia is still well into superhuman territory, with an estimated CCRL elo of 3100-3150; she'll lose badly to top engines, but will still crush any human."
Thanks for the kind words!
She's absolutely superhuman, I can guarantee you that, however I wouldn't be surprised if she sometimes had a propensity to mess up with some ridiculous sacrifice against a weaker engine.
Do you think you’ll make it multi threaded in the future? (I can only set to 1 thread)
Rebel wrote: ↑Wed Apr 03, 2024 10:57 am
The EAS tool is great if you understand its limitations. It's not perfect but it is very useful if you want to tune your (NNUE) engine.
Thank you, Ed. An honor, that the GOAT of computerchess likes my little tool...
Of course, the EAS-Tool has it's limitations - we just talked about a wrong queen-sac detection in a Rebel game right here in the forum...
The problem is, that EAS-Tool (and my other tools) must be very fast, otherwise the huge amount of enginegames in these days would be not computable. And, additionally, the EAS-Tool needs a lot of games for a valid score. So, all the work in the my tools must be done with the brutally fast pgn-parser pgn-extract. Doing own evaluations in the EAS-tool or letting the EAS-Tool just look into the evals/comments of the played games, would be way to slow. Mention, my EAS-Tool right now already needs around 1 hour to evaluate my full UHO-ratinglist pgn-file (right now it contains 48 engines and 574000 games) on a fast PC-core (4GHz).
Werewolf wrote: ↑Wed Apr 03, 2024 11:49 pm
What an incredibly fun engine!
I'm not sure about this statement though as it got into a lot of trouble against a 2300 Elo engine: "It's worth noting that Patricia is still well into superhuman territory, with an estimated CCRL elo of 3100-3150; she'll lose badly to top engines, but will still crush any human."
Thanks for the kind words!
She's absolutely superhuman, I can guarantee you that, however I wouldn't be surprised if she sometimes had a propensity to mess up with some ridiculous sacrifice against a weaker engine.
Do you think you’ll make it multi threaded in the future? (I can only set to 1 thread)
Yes, multithreading is definitely planned.
What we definitly need is a skill-level parameter in Patricia for making the engine playing weaker. Then this engine would be very cool for playing as a human against it (perhaps on a electronic chessboard, using the Android-Binary on a Smartphone). Right now, Patricia is way to strong for using it as an opponent for humans.
Werewolf wrote: ↑Wed Apr 03, 2024 11:49 pm
What an incredibly fun engine!
I'm not sure about this statement though as it got into a lot of trouble against a 2300 Elo engine: "It's worth noting that Patricia is still well into superhuman territory, with an estimated CCRL elo of 3100-3150; she'll lose badly to top engines, but will still crush any human."
Thanks for the kind words!
She's absolutely superhuman, I can guarantee you that, however I wouldn't be surprised if she sometimes had a propensity to mess up with some ridiculous sacrifice against a weaker engine.
Do you think you’ll make it multi threaded in the future? (I can only set to 1 thread)
Yes, multithreading is definitely planned.
What we definitly need is a skill-level parameter in Patricia for making the engine playing weaker. Then this engine would be very cool for playing as a human against it (perhaps on a electronic chessboard, using the Android-Binary on a Smartphone). Right now, Patricia is way to strong for using it as an opponent for humans.
That is planned as well! I think a good way to go about doing it is instead of just having it make random blunders, try and determine what sort of blunder a "human" would make - this requires having Patricia make "non obvious" mistakes (for example, it won't hang a piece, but it might hang a tactic), which will be fun to try and implement