I recently implemented contempt in Winter, which is probably the last major feature I will be adding before releasing Winter 0.7. Unfortunately I have unexpectedly measured a regression in Winter with my implementation at 0 contempt. The good news is I think I understand what might be going on, so I will hopefully be able to fix it and release Winter without compromise.
Due to how I implement contempt, it very naturally means Winter will natively support the Armageddon format. The contempt works by changing the value for a draw, which Winter can do because the development version's evaluation function estimates win, draw and loss probabilities. So I will be adding a switch, which if true sets contempt to 100 with white and -100 with black. I am unsure if this will lead to actual improvements in Armageddon play, but it at least makes a lot of sense from a theoretical perspective.
The regression has gotten me to start thinking about the subjective value of engine features. I was wondering what features people are interested in that they would not mind taking a penalty in playing strength for? I am talking specifically in the context of engines outside of the very best, as I can understand that above a certain engine ranking the primary consideration must be playing strength.
So far I have come up with a few ideas.
Features in regards to playing style, especially if the style can be easily modified such as in Rodent, are for sure worth some rating points for people who run their own private tournaments or like to play against engines themselves.
Engines good at solving chess problems are useful tools. The most prominent engine in this regard is probably Sting.
Engines giving clean evals are often more pleasant to analyze with. It is a great quality of life if an engine understands h-pawn with the wrong colored bishop will not win, regardless of how much shuffeling is done. Often times such features will have neutral or even negative Elo effect due to overhead and how rarely they change the outcome of the position, but it feels better. I think this is one reason fans really want EGTB support, they feel there is a big difference, even if Elo measurements show otherwise.
jorose wrote: ↑Sat Sep 07, 2019 10:15 pm
I recently implemented contempt in Winter, which is probably the last major feature I will be adding before releasing Winter 0.7. Unfortunately I have unexpectedly measured a regression in Winter with my implementation at 0 contempt. The good news is I think I understand what might be going on, so I will hopefully be able to fix it and release Winter without compromise.
Due to how I implement contempt, it very naturally means Winter will natively support the Armageddon format. The contempt works by changing the value for a draw, which Winter can do because the development version's evaluation function estimates win, draw and loss probabilities. So I will be adding a switch, which if true sets contempt to 100 with white and -100 with black. I am unsure if this will lead to actual improvements in Armageddon play, but it at least makes a lot of sense from a theoretical perspective.
The regression has gotten me to start thinking about the subjective value of engine features. I was wondering what features people are interested in that they would not mind taking a penalty in playing strength for? I am talking specifically in the context of engines outside of the very best, as I can understand that above a certain engine ranking the primary consideration must be playing strength.
So far I have come up with a few ideas.
Features in regards to playing style, especially if the style can be easily modified such as in Rodent, are for sure worth some rating points for people who run their own private tournaments or like to play against engines themselves.
Engines good at solving chess problems are useful tools. The most prominent engine in this regard is probably Sting.
Engines giving clean evals are often more pleasant to analyze with. It is a great quality of life if an engine understands h-pawn with the wrong colored bishop will not win, regardless of how much shuffeling is done. Often times such features will have neutral or even negative Elo effect due to overhead and how rarely they change the outcome of the position, but it feels better. I think this is one reason fans really want EGTB support, they feel there is a big difference, even if Elo measurements show otherwise.
Added to those interesting features is UCI_Elo even in a range of 1000 to 2000 then try to simulate human players.
Even better perhaps is UCI_Glicko, UCI_Glicko_LimitStrength which I have plan to implement. The advantage is that Lichess uses Glicko and you can run your engine in Lichess and apply filter to play only against humans. You will get games to study to try to simulate human play.
jorose wrote: ↑Sat Sep 07, 2019 10:15 pm
Due to how I implement contempt, it very naturally means Winter will natively support the Armageddon format. The contempt works by changing the value for a draw, which Winter can do because the development version's evaluation function estimates win, draw and loss probabilities. So I will be adding a switch, which if true sets contempt to 100 with white and -100 with black. I am unsure if this will lead to actual improvements in Armageddon play, but it at least makes a lot of sense from a theoretical perspective.
Very cool feature! I hope, other engines will follow, supporting Armageddon in their draw-eval...
