Hnefatafl forum

Ytreza wrote: Sat May 09, 2026 5:22 pm I don't quite understand how the engine can win as black in the training if it doesn't know how to surround.

In training, it wins either by surrounding the king on four sides or by preventing the opponent from making any moves.
But I haven't seen a complete victories by surrounding of all pieces.

Usually white will avoid being surrounded by placing pieces along the edges, such that black will have to win by stalemate. It's the only way to win if white plays "maximally annoying" in the endgame. So it's not surprising that this is most frequent. However it also requires that black progressively closes around the edge(s) occupied by white, which requires a similar strategy of slowly enclosing, or moving a line across the board.

Might be that the quality of endgame is exceptionnally bad both for white and black at the moment. Might also be that white makes the mistake of repeating in this kind of situation where it's stuck, which would result in an easy win for black, and acustom the engine playing black of just waiting for white to repeat instead of proactively moving toward the win.

Maybe too
I'm not accepting wins by repetitions in training to have more clean data

You mean that you force white to do something else?

Ytreza wrote: Sun May 10, 2026 9:26 am You mean that you force white to do something else?

No, I skip them and don't save them at all to reduce noise

But then the engine never learns that it's bad to repeat, no?
And repetitions might be frequent in endgame where the engine has no clue what to do.

Ytreza wrote: Sun May 10, 2026 10:38 am But then the engine never learns that it's bad to repeat, no?
And repetitions might be frequent in endgame where the engine has no clue what to do.

There may be a side effect, but it's difficult to extract any information from repetitions and understand what exactly the repetition resulted from. Anyway engine will do repetitions if it needed, because we have search.

Of course we can run experiments here and compare is it better to have repetitions in train games or not. Maybe it's not a big problem and we can keep this positions too.

I'm going to train a small network for Copenhagen later, maybe next week.
Since it looks like it won't be any weaker than the current one.

At least, it reveals the historical version very well.

As an example it took me only one try to trick it into repeating at a corner (and therefore losing despite the obvious potential - I didn't play very well as black here).

1. d1-d3 e5-e2
2. d3-d2 f8-c8
3. a4-c4 g7-g10
4. h11-h10 f7-j7
5. d11-d9 g5-j5
6. k8-j8 f6-f9
7. j8-j9 f9-i9
8. k4-i4 i9-i11
9. j9-j11 i11-i10
10. j11-j10 i10-i11
11. j10-j11 i11-i10
12. j11-j10 g6-g10
13. d9-g9 i10-i11
14. j10-j11 i11-i9
15. j11-j9 i9-i11
16. j9-j11 i11-i9
17. j11-j9 h6-h11
18. k7-k9 i9-i11
19. j9-j11 i11-i9
20. j11-j9 i9-i11
21. j9-j11 i11-i10
22. j11-j10 i10-i9
23. j10-j9

Only with 1s thinking though. With 5s it doesn't repeat.

Interesting
I've actually seen this problem with larger time controls, I'll have to look into it.