Scan's 4x4 pattern concept and non losing game

[Catherine]

Hi all programmers,
I saw like you all the DXP games played by Scan against Kingsrow and Dragon, the only and surprising thing that we notice, is that it don't lose a single game.
My question is, perhaps that the question has been approached under another angle,: that is it mean that Scan play perfectly?
I said this, because if we take into account the principle that, if the two parts play perfectly, the only result should be a draw!!!
The last but not least thing is that, like Fabien said, he programmed Scan by taking into account the fact that Chess are 90% of tactical and 10% of strategy.
It's clear that it is the same thing with draughts. The endgames of Scan prove it.
All this to ask if it is this concept of 4x4 pattern that give this strenght? I also ask to you, and it will be so kind to explain it with an example to permit to us to apply it in our game.
For me, the programmers are the new masters of the draughts community. Because if a programmer is able to program an new idea that is so strong to beat the best player of the world, i thing that by TRYING to apply also this idea, our game will be better.
Thank again to all the programmers of this forum, God bless you all.
Catherine.

[Fabien Letouzey]

Hi Catherine,

I saw like you all the DXP games played by Scan against Kingsrow and Dragon, the only and surprising thing that we notice, is that it don't lose a single game.
My question is, perhaps that the question has been approached under another angle,: that is it mean that Scan play perfectly?
I said this, because if we take into account the principle that, if the two parts play perfectly, the only result should be a draw!!!

Scan does not play perfectly. It does lose games, just less often than other programs. Of course that's the same as saying that it's closer to perfection, but we don't know how far and it's difficult to find out.

The last but not least thing is that, like Fabien said, he programmed Scan by taking into account the fact that Chess are 90% of tactical and 10% of strategy.
It's clear that it is the same thing with draughts. The endgames of Scan prove it.

I put draughts into the "positional game" category together with Othello. One characteristic is that when you are in a bad situation and you don't see any quick way out (in few moves), then most of the time there is no way out at all. This is not true in chess and similar games; deep tactics are always present. If I used Scan's pruning in chess, the program would search extremely deeply but make big tactical mistakes every so often.

All this to ask if it is this concept of 4x4 pattern that give this strength? I also ask to you, and it will be so kind to explain it with an example to permit to us to apply it in our game.

I think it is futile to try to pinpoint a single source of strength. I speak from experience in other games, and we also know that Dragon's evaluation is better than Scan's. In my opinion a program is only as strong as its weakest link, so locating weaknesses in other programs seems easier.

How to apply patterns in your own play? The theory is that you shouldn't need to do anything special. That's because Scan mimicks the human eye to some extent: it has a local vision of the board. Let's translate Scan's construction to the human world: I would say that it looked at many games by amateurs who are strong only in tactics, and discovered which shapes are good or bad (led to wins and losses respectively in those games).

So in theory, you "only" need to look at many games. In practice I don't think it works, or it takes too much time (grandmaster's life-long study). Game teachers often say that watching is useless and you need to play instead. I know Go better and there is a saying: to become one dan (black belt), you need to lose 1000 games.

Fabien.

[Catherine]

Hi Fabien,
Thank for taking time to answer me. I understood now. So to do a summary, Scan use self learning to improve his game.
Something isn't clear in my mind, you said that Dragon evaluation it's better than Scan's, but from where coming this difference of strenght ?
Thank again.
Catherine.

[Fabien Letouzey]

So to do a summary, Scan use self learning to improve his game.

Yes. Just to clarify "improve" though: Scan started with no positional knowledge at all: only "1 king = 3 men". And then as a second step it was already at GM level; it wasn't a gradual improvement as for humans. But of course, even the first version benefitted from the computing power (and accuracy) of ... computers.

Something isn't clear in my mind, you said that Dragon evaluation it's better than Scan's, but from where coming this difference of strength ?

I asked the same question for the same reason. We don't have the answer yet (only Michel can answer, and it's not easy), but here are a few possibilities, none of which seems particularly likely:
- Dragon has search bugs; as a matter of fact, Michel found one but it wasn't huge
- Dragon has the best eval but does not make full use of it; it seems logical but I don't feel that it can explain everything
- Scan has several small improvements scattered around the place (or is better tuned, which is about the same); but Michel tested changes very seriously so, again, I don't think that's the reason

I have a feeling that "all of the above" is the final answer, each of them in a small dose ...

[MichelG]

asked the same question for the same reason. We don't have the answer yet (only Michel can answer, and it's not easy), but here are a few possibilities, none of which seems particularly likely:

I have been investigating for a while now, and i still don't have the answer to that.

I am investigating all parts of the dragon source code, and trying some of the concepts that scan uses.

This has turned up some small bugs in dragon, but i also tried some ideas that scan uses. For example, scan extends the search when it is a forced move only. Dragon used to extend on all capture moves.

The difference is subtle, but changing things like this improves performance by minute amounts. It takes rigorous testing to determine whether such a change works, but all these changes add up.

With the bugfixes and changes, the current (unreleased) version has definitely improved compared to 4.5.0. At very short time settings (12 sec/game) the last damexchange match ended in +9 -9 =60, but at higher time settings scan still dominates by a large margin.

I have some ideas about this, but research continues...

Michel

[Catherine]

Ok, i beginning to understand. It's most a question of concept.
Michel said that compared to Dragon which extand search when there's moves capture, Scan used to extand it's when there's forced moves; a concept that we can apply too.
I think that with this idea, the search become so deep and more logical.
Because the end of the forced moves will show more cleary the probably true result
Win, lose or draw.
Fabien said also that Scan it's more stable and better computed.
It's a very important point to fix in any program.
What news about Damage and Kingsrow ?
Thank for all.
Catherine.