chess tiger on iphone

[ozziejoe]

the last thread related to chess tiger seemed to get hijacked by clone talk so i'd thought i'd start a new one about chess tiger.


I think this is a beautiful program . Everything works simply, the interface is clear and pleasing, the help is easy to access. I think this is a shining example of what an Iphone app should look like and do.

The training levels are a real pleasure. The weaker levels do play a bit like weaker humans I play against (although the weaker levels do seem to be blind to simple, long term plans such as advancing a pawn to queen.)

There is something i would love to see in a future version. When playing against the weaker opponent, randomly vary the playing style and keep that style secret from the player. So I should not know if i am going up against a normal, aggressive, or super aggressive player.

There are a few styles, in addition to the above three, that would be easy to implement and are common amongst average to weak players

1) the imitator (starts out by imitating your moves)

2) the sniper (called hypermodern by some. fianchettoed bishops,etc; this is merely a book preference, but the preference should be there no matter what the opponent plays, So maybe you could have the double fienchetto every time

3)The fortress builder (trys to set up impregnable fortress, e.g., stonewall..again should require this personality to play this style against every opening)

4) the king chaser (i assume this is sorted by your aggressive, very aggressive , and suicidal mode. Definitely the most common type of bias in under 1800 players. Indeed, I've built an entire opening system to deal with it).

5) the offbeat opening player (again, just a book preference. the main thing I am proposing is you don't know what you are going to get when you play)



Anyway, just some thoughts on improving an already excellent program.

And as for whether it can beat rybka or one of the other cut-and-paste programs, who cares? Honestly, do we need any more software for analyses. Surely, we have enough. we need engines that we enjoy playing.

best
Joseph

[ozziejoe]

One thing i've disliked about previous programs that try to play weak is that they are so variable in their strength. so an engine set to play a 1700 level will play like an absolute moron and then like a grand master...totally unrealistic stuff.

I just played the 1750 level of chess tiger, and had to find true tactical shots to beat it. It played consistently at a particular level, not suddenly becoming strong and then weak

The important thing here is that chess tiger allowed the tactical shot. some of the other engines, on weak level, will just give you pieces but rarely let you have a tactical shot.

Christophe, if you have time, maybe you can say something about how you designed the weaker levels and how you bench marked them? Does a 1750 player win about half the time against your 1750 setting? How did you design the program so it seems to lose in a realistic way.

In all this rush to make programs that win all the time, christophe has designed a program that loses beautifully!

best
Joseph

[ozziejoe]

I spoke too soon. Chess tiger just gave away a knight on the 1910 (grade A) training level, making a move that no grade a player would ever make.

If you want a realistic game from a chess game, then i think some processing resources need to be devoted to playing humans (resources that do not increase performance). For example, if you have a quad processor, you could use one of the processors for the sole purpose of creating tactical shots in a position. e.g., at an easy level , it would create simple forking tactics. at a harder level, it would create tactics that require two moves and sac of material to work. etc....

So to be clear...I want an engine that loses in a way that a human would lose..chess tiger is close....I don't think chessmaster is very realistic because it has that tendency to play super-weak and then suoer strong.

best
J

[tiger]

I spoke too soon. Chess tiger just gave away a knight on the 1910 (grade A) training level, making a move that no grade a player would ever make.

If you want a realistic game from a chess game, then i think some processing resources need to be devoted to playing humans (resources that do not increase performance). For example, if you have a quad processor, you could use one of the processors for the sole purpose of creating tactical shots in a position. e.g., at an easy level , it would create simple forking tactics. at a harder level, it would create tactics that require two moves and sac of material to work. etc....

So to be clear...I want an engine that loses in a way that a human would lose..chess tiger is close....I don't think chessmaster is very realistic because it has that tendency to play super-weak and then suoer strong.

best
J

I would be curious to see the position where Chess Tiger blundered.

It happened to a friend of mine during the beta test stage of the iOS version. He showed me the position and I ran it in the Linux version, which displays more information.

We realized that Chess Tiger had actually attempted a tactical sacrifice, but made a mistake (an intentional one) in computing the variants. The opponent could lose if it did not answer correctly. Unfortunately the correct answer was quite obvious so it did not even look like a sacrifice. But it was.

So my friend was still holding to his opinion that the engine blundered, but changed his opinion about the fact that a human player could not have done it.

Do you still have the game?


I'm answering here to your question about how the trainer levels have been designed:

In previous versions of Chess Tiger, and in particular in the PalmOS version, the trainer levels were simply depth-limited levels. This is a very crude way of weakening the program and it was not very useful after the first 3 or 4 levels (it was quickly getting much too strong).

I rewrote all of that for the iOS version. Now each trainer level is defined by 4 adjustable parameters, 2 parameters defining how clever the engine is in various aspects and a third parameter which controls randomness in order to get an even finer setting. There is also a parameter controlling the depth of the opening knowledge of the level.

After implementing those parameters, I thought that it would have been fairly easy to create levels spanning regularly from 850 to 2100 FIDE elo... It has not been at all.

I had to invest TWO WEEKS to adjust the parameters for these 28 levels.

I dedicated two 2.4GHz Quads to the tests. After adjusting one level, I played 25000 games against the two levels below or the two levels above (so that makes 50000 games for each test).

So if I get the parameters for one level correct at the first try, it takes 50000 games to confirm that the level is correctly adjusted. For 28 trainer levels correctly adjusted at the first try it is necessary to play 24*50000=1.2 million games (the first 2 levels are excluded as they are special).

Naturally my trainer levels were absolutely not adjusted correctly, and finding the right parameters for one level, based on the previous two for example, took a number of tries (typically between 4 to 7 because I always ran the matches a second time when the parameters looked right).

I have not kept track of all the test matches but I think I had to let the computers play a grand total or 5 to 7 million games. The games were extremely quick because the computational effort to compute a trainer level move is both fixed and quite modest, but I had to wait for each match to finish, write down the result and then decide how to adjust the parameters for the next match.

It took 2 very boring weeks.

On the FIDE elo scale, I think my levels are equally spaced. I think the difference between a level and the next one is 50 elo points, but it could be argued that this number is not valid because the engine played against itself. The difference in strength may not apply to other opponents.

I think that at least these levels are evenly spaced.

I cannot say with any certainty that they are exactly the elo strength they claim to be. I have not been able to get enough games against rated human players to be sure.

I have done my best to be as rigourous as possible in establishing these levels, but the scale could need some adjustment with the help and experience of rated players. By adjustement I mean that the trainer levels could have to be shifted (a constant elo should be added or substracted to all trainer levels) or even scaled (if it is shown that the difference between each consecutive level is not 50 elo points).

This could be done by establishing with a good number of games the performance of the program against beginners around 850 elo points and against expert players around 2000 elo points (2100 would be even better).

The most important fact is, I think, that the levels are evenly spaced and close enough to each other that you can find a level that approximately matches your own level and that gives you a good reference for keeping track of your progress.

I guess these two weeks were worth it after all. It's better to invest two weeks in order to provide the users with reliable, progressive levels than to improve from 2700 elo to 2720 elo, which is meaningless for 99.99% of my users.


// Christophe

[ozziejoe]

Wow, you've gone through a lot of trouble. It has payed off in the ease in which the player can find a level that perfectly suits him/her

Thanks to the really cool function of being able to load your previous game, I was able to reload the game with the blunder. This is on iphone and chess tiger is set to 1910 ELO. I was playing pretty fast so this game is not a work of art (and chess tiger was playing faster than me)

Chess tiger has just played NG4, which looks to be a decent size blunder (evaluation goes from .12 to over +4. This move does give some interesting attacking chances on the H file, but this is probably too obvious a blunder for a 1900 player.


1.Nf3 c5 2.c4 Nf6 3.g3 b6 4.Bg2 Bb7 5.O-O g6 6.Nc3 Bg7 7.d3 Ng4
8.Bd2 h6 9.Rb1 d6 10.a3 Qd7 11.b4 Qf5 12.bxc5 Qxc5 13.Qb3 Nd7
14.h3 Ngf6 15.Be3 Qa5 16.Bd2 Nc5 17.Qa2 Qa6 18.Nh4 Ng4 *



[d]r3k2r/pb2ppb1/qp1p2pp/2n5/2P3nN/P1NP2PP/Q2BPPB1/1R3RK1 w kq - 9 19[/d]

[tiger]

Wow, you've gone through a lot of trouble. It has payed off in the ease in which the player can find a level that perfectly suits him/her

Thanks to the really cool function of being able to load your previous game, I was able to reload the game with the blunder. This is on iphone and chess tiger is set to 1910 ELO. I was playing pretty fast so this game is not a work of art (and chess tiger was playing faster than me)

Chess tiger has just played NG4, which looks to be a decent size blunder (evaluation goes from .12 to over +4. This move does give some interesting attacking chances on the H file, but this is probably too obvious a blunder for a 1900 player.


1.Nf3 c5 2.c4 Nf6 3.g3 b6 4.Bg2 Bb7 5.O-O g6 6.Nc3 Bg7 7.d3 Ng4
8.Bd2 h6 9.Rb1 d6 10.a3 Qd7 11.b4 Qf5 12.bxc5 Qxc5 13.Qb3 Nd7
14.h3 Ngf6 15.Be3 Qa5 16.Bd2 Nc5 17.Qa2 Qa6 18.Nh4 Ng4 *



[d]r3k2r/pb2ppb1/qp1p2pp/2n5/2P3nN/P1NP2PP/Q2BPPB1/1R3RK1 w kq - 9 19[/d]

OK, no excuse here, this one is a failure.

But I see why and maybe I can improve this. Thanks for posting the position, it gives me food for thought.


// Christophe

[Werner]

Hi,
is it possible to run Chess Tiger app on a IPhone emulator from genuitec at normal speed?

Happy Easter

Werner

[tiger]

Hi,
is it possible to run Chess Tiger app on a IPhone emulator from genuitec at normal speed?

Happy Easter

Werner

No sorry.

What Genuitec provides is NOT an iPhone emulator. It's one of the most misleading advertisement I have ever seen.

With their tools you create apps using HTML5, JavaScript and CSS3, and they will run inside the Safari browser on the iPhone.

They will also run on other browsers compatible with these technologies.

In no way do they allow to run an iPhone app on Windows, what a real emulator is supposed to do.


// Christophe

[ozziejoe]

I am curious Christophe. What did chess tiger see in this position that made it want to sacrifice. It did lead to some dangerous threats on the h file and some potential knockout shots featuring the rook and queen focusing on h1

[tiger]

I am curious Christophe. What did chess tiger see in this position that made it want to sacrifice. It did lead to some dangerous threats on the h file and some potential knockout shots featuring the rook and queen focusing on h1

Actually this has been caused by the "handicaped" search.

I play with the horizon effect in order to create distant tactical mistakes that you can take advantage of if you can see deep enough.

At this level the random factor I talked about had a small chance to reduce slightly the horizon in order make the program blind to a potential combination 4 or 5 plies deep.

I think it involved Bxg2 Kxg2 Qb7+ and then Bxc3. Tiger thought you could not recapture on c3.

The problem here is that the line is quite forced, and creating an horizon effect in a forced line results in a too obvious mistake.

If you let Tiger replay its last move you will see that it may often prefer O-O instead.

So it was a combination of a too aggressive horizon effect and bad luck.

I really want to work harder on this as soon as I have some time for that. For now, I hope it does not happen too often.

To be fair, you should also count the number of blunders you make yourself... We all do fall in obvious traps from time to time, or sacrifice something just to realize there is nothing to gain.


// Christophe