The new Tafl engine (AI)

[sovaz1997]

I played against it with the blacks on the Level 21 with 1s per move and managed to win (I also played blitz, thinking maybe 1-3 seconds per move). I am definitely stronger than level 21 (1s per move). However, I am sure it is stronger than me, if it has more time to think. I also recognized that it didn't detect that I won in one of the games in which I managed to take all of its pieces and blocked its king without possible move. So, the game continued despite the engine couldn't make a move.
I think the engine is strong enough to give us the evaluation of the world champions. I hope you will be able to analyze the championship games of all of the Hnefatafl world champions and present us with statistics. I wonder who played the best tafl over the years. Some interesting questions are:
How good was Schachus in 2014? Was he back then better than Casshern now?
How well I played in my peak in 2021? Was I playing better in 2021 than in 2024 when I won my second world championship?
How good was Plantagenet in 2017 when he defeated a green version of Casshern? Will Plantagenet from 2017 played better tafl than Casshern nowadays?
Did the overall strength of the players has increased over the years or Schachus, Nath and Crust were superior?

For the research, you can take only the championship games for all of the winners in the world championships and analyze the games. I think we need to exclude from the analyzes positions that are completely won by one of the sides. For example, the engine will still detect some errors when the blacks are winning by blocking all of the corners and have a black cordon build because in such a position, the blacks might not play the optimal moves but they actually don't need to find the best moves. Or when the whites are in a completely lost position, they might choose to play a sub-optimal move that has more potential to bring some complications on the board. Such positions should be excluded unless the errors led to losing the game or completely changed the chances in the position. Maybe the best way is to exclude analyzes of positions in which one of the sides have an advantage more than 50 points.

Hey
It's nice, yes, i have plans to analyze champions games!

[sovaz1997]

I played against it with the blacks on the Level 21 with 1s per move and managed to win (I also played blitz, thinking maybe 1-3 seconds per move). I am definitely stronger than level 21 (1s per move). However, I am sure it is stronger than me, if it has more time to think. I also recognized that it didn't detect that I won in one of the games in which I managed to take all of its pieces and blocked its king without possible move. So, the game continued despite the engine couldn't make a move.
I think the engine is strong enough to give us the evaluation of the world champions. I hope you will be able to analyze the championship games of all of the Hnefatafl world champions and present us with statistics. I wonder who played the best tafl over the years. Some interesting questions are:
How good was Schachus in 2014? Was he back then better than Casshern now?
How well I played in my peak in 2021? Was I playing better in 2021 than in 2024 when I won my second world championship?
How good was Plantagenet in 2017 when he defeated a green version of Casshern? Will Plantagenet from 2017 played better tafl than Casshern nowadays?
Did the overall strength of the players has increased over the years or Schachus, Nath and Crust were superior?

For the research, you can take only the championship games for all of the winners in the world championships and analyze the games. I think we need to exclude from the analyzes positions that are completely won by one of the sides. For example, the engine will still detect some errors when the blacks are winning by blocking all of the corners and have a black cordon build because in such a position, the blacks might not play the optimal moves but they actually don't need to find the best moves. Or when the whites are in a completely lost position, they might choose to play a sub-optimal move that has more potential to bring some complications on the board. Such positions should be excluded unless the errors led to losing the game or completely changed the chances in the position. Maybe the best way is to exclude analyzes of positions in which one of the sides have an advantage more than 50 points.

By the way how much speed to you have here, how much nodes per second?
It would be interesting to know what speed your engine runs at; that affects its playing strength.

[sovaz1997]

Overall, there are plans to train an even larger network, and subsequently derive a distilled, smaller version from it that will perform well on a CPU.
The historical hnefatafl and tablut also plays better, after yesterday updates, if you interested you can take a look.

[sovaz1997]

I played against it with the blacks on the Level 21 with 1s per move and managed to win (I also played blitz, thinking maybe 1-3 seconds per move). I am definitely stronger than level 21 (1s per move). However, I am sure it is stronger than me, if it has more time to think. I also recognized that it didn't detect that I won in one of the games in which I managed to take all of its pieces and blocked its king without possible move. So, the game continued despite the engine couldn't make a move.
I think the engine is strong enough to give us the evaluation of the world champions. I hope you will be able to analyze the championship games of all of the Hnefatafl world champions and present us with statistics. I wonder who played the best tafl over the years. Some interesting questions are:
How good was Schachus in 2014? Was he back then better than Casshern now?
How well I played in my peak in 2021? Was I playing better in 2021 than in 2024 when I won my second world championship?
How good was Plantagenet in 2017 when he defeated a green version of Casshern? Will Plantagenet from 2017 played better tafl than Casshern nowadays?
Did the overall strength of the players has increased over the years or Schachus, Nath and Crust were superior?

For the research, you can take only the championship games for all of the winners in the world championships and analyze the games. I think we need to exclude from the analyzes positions that are completely won by one of the sides. For example, the engine will still detect some errors when the blacks are winning by blocking all of the corners and have a black cordon build because in such a position, the blacks might not play the optimal moves but they actually don't need to find the best moves. Or when the whites are in a completely lost position, they might choose to play a sub-optimal move that has more potential to bring some complications on the board. Such positions should be excluded unless the errors led to losing the game or completely changed the chances in the position. Maybe the best way is to exclude analyzes of positions in which one of the sides have an advantage more than 50 points.

I found seems in my back-office!
Is this game your played?


From what I could see, the average speed was about 100 times slower than what I get on my GPU (7 nodes/second).

It seems your computer either doesn't have a GPU, or for some reason, it didn't activate. Anyway, I’ll prepare a CPU-based version; it will naturally be less powerful, but perhaps the difference won't be all that significant.

[Draganov]

Yes, this was the game that I won and yes my computer is slow and old. I am using a Lenovo Thinkpad W530 laptop. The speed of the engine is 24 n/s when analyzing positions.
Maybe this is the reason why I won against it

[sovaz1997]

Yes, this was the game that I won and yes my computer is slow and old. I am using a Lenovo Thinkpad W530 laptop. The speed of the engine is 24 n/s when analyzing positions.
Maybe this is the reason why I won against it

Hehe, no problem!
You could try playing against the CPU-optimized network, I'm just uploaded it to the site. It doesn't have levels, just one strongest level.

It’s a bit weaker, but it was trained on games played by the best, larger network, and it runs much faster when using the CPU.