Truus DXP-Server

[Ed Gilbert]

I played a few automated games between truus and kingsrow where kingsrow was configured to use no endgame databases at all. Although these games were draws, each game went to 75 moves and each result was logged as 'unknown' because it could not get a database score. So I made a small change to the eval function: if number of pieces is <= 4 and at least one king on each side, return the same eval score as database draw. That small change made a big difference in its ability to resolve games. I then ran a full match this way. The settings were the same as the 6pc and 4pc matches except kingsrow did not use any endgame databases. Result:

Kingsrow vs Truus: 22 wins, 2 losses, 112 draws, 22 unknowns

I was able to determine a conclusive result in all the unknowns through analysis with the 9pc db. Adjusted results:

Kingsrow vs Truus: 26 wins, 4 losses, 128 draws

So kingsrow seems to play just as well without any endgame db at all as it does using a 4pc db. In fact these results are slightly better than the 4pc match, but I'm sure if Rein runs these numbers through his statistics program it will say that 158 games is not nearly enough to distinguish a difference that small with any reasonable confidence.

-- Ed

[BertTuyt]

Ed, interesting result.
Im now working on the Ballot Generator, and the additional PDN support.
Hope to have something up and running this weekend.

Just a few questions , for curiosity:
* Around which movenumer does KingsRow get the material advantage (so 1 man ahead, or a breakthrough).
* Do you see a general pattern in the KingsRow wins, do they occur in specific situations, in the late midgame, or randomly distributed.
* How many ply do you outsearch Truus during these games.

Bert

[Ed Gilbert]

Just a few questions , for curiosity:
* Around which movenumer does KingsRow get the material advantage (so 1 man ahead, or a breakthrough).
* Do you see a general pattern in the KingsRow wins, do they occur in specific situations, in the late midgame, or randomly distributed.
* How many ply do you outsearch Truus during these games.

Bert, the match takes two days, and I watched at most 10% of the games, so it is a little hard to generalize. Truus does always seems to display a search depth which is several ply smaller than kingsrow, but this is only the nominal depth, and it does not display average leaf depth or max depth which might be more interesting. Truus also has the tactical table extensions that are described in Keetman's masters thesis, and kingsrow does not have anything like this, so it's likely truus extends further beyond its nominal depth than kingsrow. In the few decisive games that I watched I think kingsrow did start to see a big score earlier than truus.

Very often truus and kingsrow disagree on which side has the better position. We noticed the same disagreement during the damage/kingsrow matches. I don't know if kingsrow just has a very different notion of what is good, or this is normal. I have also played a large number of games against Flits, and I think Flits and kingsrow tend to agree more on positional strength. Flits also seems to be a bit stronger than truus, judging from my results against it, but not a great deal stronger. I can only tell you that it is not uncommon for truus to think that it has a great position, and at the same time kingsrow thinks its position is also quite good! In 8x8 English checkers this much disagreement between two strong programs is rare.

BTW, there is a bug in the truus GUI -- it has a GDI object leak. Every time it paints the board it consumes another GDI object. You can see this using Windows task manager. After about 75 games my computer ran out of GDI objects and the whole desktop display got very corrupted, although the programs continued to play without any problems. I stopped the match and restarted truus and then it was good for the rest of the match. You will probably run into this when you do a truus/damage match.

-- Ed

[Ed Gilbert]

if number of pieces is <= 4 and at least one king on each side, return the same eval score as database draw

I need to correct this because it's not exactly what I changed. There was already code there to return a small value and I only changed the value it returned to +/-1 which is what I use for db draws. But the actual logic is: if the max pieces on either side <= 3 and at least one king on each side, then reurn db draw score. So it also handles 3x2, and 3x3 cases.

-- Ed

[BertTuyt]

Today Damage played again a testgame against Truus, with a strange win for Damage.
In a position which was better for Damage but defensible, Truus played a strange move which quickly leaded to a breakthrough.
The strange thing that the scores of Truus at that point did not reveal the problems "she" was facing, whereas the Damage score already revealed the catastrophe for Truus.

When i analysed the game later, Truus also immedialtely found the mistake (red line in the score graph), but that could be related to the backwards search process.

But also when i set up the position then Truus plays from the start the good/better move, without ever considering the fail-move.

I don't know if this is related to the DXP-server, or that there is a bug in Truus. But something is wrong here.
Maybe also others have seen strange behavior of Truus near the late middlegame whith strange scores or moves, so far I don't have any explanation yet.

I will study the fatal position again, and if i cant get a clue, I will post it for others to examine.

Bert

[Ed Gilbert]

Hi Bert,

While I haven't seen truus make any really terrible blunders, I have noticed something about the way it manages its clock. It sometimes uses a large portion of its game time on a few of the early moves, and then after about 30 or so moves it starts making some moves quickly after only a shallow search. I think I have seen it get into trouble as a result of this.

If you try to reproduce the move, remember that it uses pondering, so to more closely reproduce the situation from your game you should set up the predecessor position where it was damage's move and let truus ponder for a few seconds, then make damage's move and see what truus does.

I would be very surprised if the dam exchange has any effect on truus' search, as in its final form the dxp server simply pokes ascii commands into the truus engine command buffer, just as the truus GUI does. It does not directly modify any engine variables used during the search. I also found that its strength in dam exchange games is similar to the strength I found in over 100 games played manually. The clock management issue is also there in manual games.

The only other engine I have access to that has dam exchange is Dam2.2. Maybe I will set up a match between dam and truus and see how that goes.

-- Ed

[BertTuyt]

I replayed the game but it was difficult (= impossible) to get the same move back from Truus.
Think Ed was right, it is a complex combination of time control and hash-tables.
Anyway would like to share the diagram with you all.
In the below position (white to move) Truus played 39-34, and eventually lost.

[img]http://fmjd.org/dias2/save/12377319635.png[/img]

Bert

[FeikeBoomstra]

After 39-34 isn't it a draw?

1.39-34 40-44 2.34x23 13-18 3.50x39 18x29 4.28-22 17x28
5.38-33 29x38 6.49-43 38x49 7.39-34 49x21 8.26x 8 45-50

[BertTuyt]

Feike, i made an error, this is the right diagram.
The white man should be place on 48 in stead of 49.
This (below) is the right diagram.

[img]http://fmjd.org/dias2/save/12377422707.png[/img]

Bert

[FeikeBoomstra]

1.39-34 40-44 2.34x23 35-40 3.50x39 45-50 4.39-33 13-18 5.16-11 18x29
6.11x22 12-18 7.22x13 40-45 8.33x24 50x 9 9.26-21 9x31 10.24-20 45-50
11.48-42 31x48 =

[BertTuyt]

Feike , why the 4. 13-18 and not

4. 40-44 5. 48-43 50-45 6. ...............................

Bert

[FeikeBoomstra]

you are right, at all search depth Horizon returned 40-44 except for the last one I looked upon, that returned 18-13, the next depth returned again 40-44 with a better score.

[BertTuyt]

TDXPS, episode IV

In this new episode (a very short one, a cliffhanger so to say), we focus on the question "Does Truus has a secret backdoor", and the related questions "How can you activate this" , and "What new opportunities and possibillities do you get?".

For now we only limit this topic with an affirmative YES !
Yes, Truus has a secret backdoor, and yes it is possible to activate it.

To be continued.............

Bert

[FeikeBoomstra]

And it is not one of things like aggressiveness?
My youngest son like to find all these cheats in computer games, but he just find them on internet, not by his own investigations.

[BertTuyt]

Feike, what do you mean?
It is not aggression, with the open Truus Backdoor you can activate very interesting options which are not provided through the standard menu.

Bert