search

Adam-Kulju / Patricia

Super aggressive chess engine
MIT License
72 stars 9 forks source link

a strangely lost game #3 - plummeting evaluation #98

Open caissist opened 3 weeks ago

caissist commented 3 weeks ago

Hi Adam, this time Patricia ver Aug 06, 2024 was playing! It looks to me like both games are pretty similar in terms of engine behavior?!

If my impression is correct, then there is a serious problem... Patricia game - lost game #03.txt Patricia game - lost game - logging #03.txt

Adam-Kulju commented 3 weeks ago

The blunders/leaps in evaluation here are only a couple pawns at a time, which is much more in line with what low skill level Patricia should do than the first games - what settings were you running her with?

caissist commented 3 weeks ago

The blunders/leaps in evaluation here are only a couple pawns at a time, which is much more in line

So what do you think was happening here: was this game lost by Patricia based on its playing skills, or based on a bad opening line (book)?

Both versions Sep 18, 2024 and Aug 06, 2024 were/are playing with these settings: single threaded, hash 1024/2048 Mb, MultiPV 1. EDIT: UCI_LimitStrength false

By the way: what do you think is correct in single threaded mode - hash 512/1024/2048 ?

Thank you.

Adam-Kulju commented 3 weeks ago

very weird. The log has it performing multiPV searches:

19:41:16.022<-info multipv 1 depth 1 seldepth 1 score cp 77 nodes 30 nps 2500 time 12 pv e5e4 
19:41:16.023<-info multipv 2 depth 1 seldepth 1 score cp 52 nodes 64 nps 4923 time 13 pv e8g8 
19:41:16.023<-info multipv 3 depth 1 seldepth 1 score cp 48 nodes 96 nps 7384 time 13 pv d8e7 
19:41:16.024<-info multipv 4 depth 1 seldepth 1 score cp 38 nodes 126 nps 9692 time 13 pv a7a5 
19:41:16.024<-info multipv 5 depth 1 seldepth 1 score cp 35 nodes 155 nps 11923 time 13 pv h7h6

but I can't replicate that behavior when I boot up Patricia and send in those commands myself. Was the Skill_Level or UCI_Elo option turned on or anything?

I'd use a hash of 512 mbs for single threaded play, maybe even lower depending on the time control. CCRL uses 256 mbs of hash for example.

caissist commented 3 weeks ago

Here a complete log file for a single game. You see all engine settings which are wanted so far. And again the engine is performing multiPV searches?! EDIT: and engine option MultiPV is set, too!? Patricia game - engine settings - logging.txt

Was the Skill_Level or UCI_Elo option turned on

Please, what does that mean?? To my understanding Skill_Level and UCI_Elo options will be activated only if UCI_LimitStrength is set to true?! It's a simple connection that most people certainly understand. And on Sep 28th, 2024 I called all the engine options I used, including "UCI_LimitStrength false"...

I'd use a hash of 512 mbs for single threaded play, maybe even lower depending on the time control.

Almost all time controls are Fisher clocks 5 1. Are 512 Mb correct?

Adam-Kulju commented 3 weeks ago

It's because you set the UCI_Elo option: UCI_Elo value 3000

On Sun, Sep 29, 2024 at 12:18 PM caissist @.***> wrote:

Here a complete log file for a single game. You see all engine settings which are wanted so far. And again the engine is performing multiPV searches?! Patricia game - engine settings - logging.txt https://github.com/user-attachments/files/17180131/Patricia.game.-.engine.settings.-.logging.txt

I'd use a hash of 512 mbs for single threaded play, maybe even lower depending on the time control.

Almost all time controls are Fisher clocks 5 1. Are 512 Mb correct?

— Reply to this email directly, view it on GitHub https://github.com/Adam-Kulju/Patricia/issues/98#issuecomment-2381432609, or unsubscribe https://github.com/notifications/unsubscribe-auth/AUP3UBX7EV5DA2UXSOEUHQDZZAY7JAVCNFSM6AAAAABO7X4FYSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGOBRGQZTENRQHE . You are receiving this because you commented.Message ID: @.***>

caissist commented 3 weeks ago

It's because you set the UCI_Elo option: UCI_Elo value 3000

Adam, that's way too short - not sure if I understood correctly. UCI_Elo value 3000 is a default engine option?! But I set "setoption name UCI_LimitStrength value false" so UCI_Elo value 3000 should be correct.

Adam-Kulju commented 3 weeks ago

hi, sorry for the late reply. if you set the UCI_Elo value or the Skill_Level value, the engine will treat it as if UCI_LimitStrength is true. I did it properly before but changed it because people kept trying to set the skill level using only UCI_Elo and wondering why it didn't.

caissist commented 3 weeks ago

Adam, again way too short. The Skill_Level and UCI_Elo options are defaulted to the highest level, and setting also UCI_LimitStrength value to false should do it without any doubt - namely playing on highest level, and nothing else?!

people kept trying to set the skill level using only UCI_Elo That is the point - me, too! To my understanding there is a need for only 2 options:

  1. UCI_LimitStrength. When set to "false" no need for anything else, all is already said. And only when setting to "true" then a 2nd option is necessary, namely:
  2. Skill_Level or UCI_Elo - either or!

This is a simple logic everybody can understand...

Adam-Kulju commented 3 weeks ago

It's because you set the UCI_Elo option: UCI_Elo value 3000

Adam, that's way too short - not sure if I understood correctly. UCI_Elo value 3000 is a default engine option?! But I set "setoption name UCI_LimitStrength value false" so UCI_Elo value 3000 should be correct.

By default it doesn't expect to be sent in any "setoption name UCI_Elo value x" commands. I will fix this for Patricia 4. (I have more bugs than I can shake a stick at it seems!)