WBEC Ridderkerk new results.

[Leo Dijksman]

Geschrieben von:/Posted by: Leo Dijksman at 05 April 2004 23:34:08:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Sune Fischer at 05 April 2004 22:34:24:

Version 149 has the same threaded design as 151 and 155, so it will have
the same bug.
Keep the new books I'm sure they are better.
They are a little smaller so engines with learning might
be able to repeat won games like Amateur did.
I just hope not too many engines in that division has book learning.

Take your time, Frenzee will not play the next 2 days!

Ok, I'll cross my fingers then

Thats strange, I cannot remember I had problems with 149 here:
Promo tourney C: - First 3 to the 2nd div, 4, 5 and 6 to the 3rd division!
ATHLON-MP2200, 2004.01.22 - 2004.01.28
Score Fren Terr Cere Esc Knig Butc
------------------------------------------------------------
1: Frenzee 149 13.0 / 20 XXXX 0110 =11= 10=1 =110 0111
2: Terra 3.3b4 13.0 / 20 1001 XXXX 0=01 1111 1100 111=
3: Cerebro 1.23a 10.0 / 20 =00= 1=10 XXXX 0=01 0=11 =110
4: Esc 1.16 9.5 / 20 01=0 0000 1=10 XXXX 11== =011
5: Knightx 1.81 9.0 / 20 =001 0011 1=00 00== XXXX 0111
6: Butcher 1.42c 5.5 / 20 1000 000= =001 =100 1000 XXXX
------------------------------------------------------------
60 games: +28 =12 -20 *0

Thats the same computer as they play now.
I will look at it tomorrow!
Frenzee will need that, it will play vs Amy

It's a random bug, perhaps they weren't very good at ponder guessing frenzee's
move?
Too bad really because if we had spotted the problem back then I might
have been able to fix it in time.
It can't do much worse than it already is
-S.

Just to be sure I have start 146 and 149 playing 2 games (at 40/5, ponder=on) vs each other and 17 other opponents, I will send the result tomorrow.
The first one is Frenzee 149 vs 146 2-0 )
Best wishes,
Leo.





WBEC Ridderkerk homepage.

[Peter Fendrich]

Geschrieben von:/Posted by: Peter Fendrich at 05 April 2004 23:40:18:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Sune Fischer at 05 April 2004 23:04:56:

A 50% time penalty is 70 Elo (according to many estimates). There are some indications that pondering actually does better than a 50% horsepower increase and especially if the ponder move guess is right 90% of the time or more.
So if your new program is not at least 70 Elo stronger than the old, it is definitely better to use the old. Maybe as much as 100 Elo.

Let's do some math here...
Engine A is pondering with 100% right guesses.
Engine B is not pondering at all.
Two processes are used and 40 moves in 40 minues.
B is using 1 minute for each move as a mean value. Let's say exactly 1 minute/move.
A is pondering 1 minute and thinking 1 minute (without sofisticated time usage).
This gives exactly doubbled time and full usage of the time alotted for both. No doubt.
The question is what could change this?
One possiblity: Let's say that they are using smart time algorithms, thinking longer on hard positions.
Could this favour A? I'm not sure but maybe. I can't see how this is better than the first case described. Well, at least A have more time to play with. It's like how richer are getting even richer because of all the money available when needed.
Anyway, if A is really earning more than doubbled time just because of having more time to play with, then doubbling the n/s in your program would give you the same advantage just by having more nodes to play with per second.
I doubt it. Do you have som figures?
/Peter
PS. The 70 ELO earned by doubbling the time is an very old SSDF observation. I'm not sure it is even true anymore and at least not for all playing levels.

I don't think that is the usual way of pondering.
Usually you can return instantly if you have already pondered longer than
what you intend to use on this move (unless if you're failing low or something).
If you have only pondered for half the time you need just keep searching until
the full time is up.
So assuming you pondered the correct move you have saved time.
In principle you can run entirely on the opponent's clock by doing it like this.
Near the end he won't have much time, so you will start to 'allocate'
relatively more time and thus not get the instant ponder hits.
But at this point you should be making better moves because you can tap into
all the time you have saved.
I'm not sure the double time would be a good simulation of this.
The question is what happens if they start off the game by using the same amount of time, but one engine saves up a lot and has maybe 3-4 times as much on the clock for the endgame?
Is this equivalent to having twice as much time in the entire game?
I don't think 70 Elo is accurate, it can depend on a lot of factors like the strength of the engines and the speed of computer or the time control.
My guess is there is diminishing returns for strong engines and long time controls.
-S.

Of course not. I just wanted an easy case to start with to clear the mind...
Later on I developed it.
I think it's best to stick to the 100% hit rate example - it's easier. That would be almost the same as having two minutes without pondering.
When B moves you have the option to continue your thinking or to stop. Exactly as if you had two minutes and no pondering.
Moving instantly is not for free. You have more time to use to get a better decision.
The 2 minute case as opposed to pondering is even slightly better because with a smart time algorithm you can sometimes move after 30 seconds if wanted. In the pondering alternative you can't move until the opponent moves.
Yes it is. Saving time to later usage is taking a risk. It's not for free compared to use the time. Game theoretical it's even wrong to not use more of the time in the beginning of the game and save it for later usage. The game could end before that.
/Peter

[Sune Fischer]

Geschrieben von:/Posted by: Sune Fischer at 05 April 2004 23:44:36:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Leo Dijksman at 05 April 2004 23:34:08:

Version 149 has the same threaded design as 151 and 155, so it will have
the same bug.

Thats strange, I cannot remember I had problems with 149 here:

It's a random bug, perhaps they weren't very good at ponder guessing frenzee's
move?

Just to be sure I have start 146 and 149 playing 2 games (at 40/5, ponder=on) vs each other and 17 other opponents, I will send the result tomorrow.
The first one is Frenzee 149 vs 146 2-0 )
Best wishes,
Leo.

Okay Leo, if you can magicly get 149 to play without hanging, then please do!

-S.

[Peter Fendrich]

Geschrieben von:/Posted by: Peter Fendrich at 06 April 2004 00:06:52:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Dann Corbit at 05 April 2004 23:19:21:

------------------------------------
The Baron 1.3.0b3 1101 3.0/4
Movei 0.08.178 0010 1.0/4
------------------------------------
Phalanx XXII 1111 4.0/4 !!
Ufim 4.04 0000 0.0/4
------------------------------------
Terra 3.3b10 ½1½1 3.0/4
Frenzee 155 ½0½0 1.0/4
------------------------------------

Of course I am beaming!


Oh my goodness, Phalanx awakens. The world may never be the same.

Terra is proceeding according to plan...

No surprises here.

Frenzee is going to do really bad, it is playing with ponder OFF.
Leo, I'm starting to think it might be better to use the old v.146 with pondering.
I think when only one engine ponders the other engine will always get into a real time jam.
-S.

A 50% time penalty is 70 Elo (according to many estimates). There are some indications that pondering actually does better than a 50% horsepower increase and especially if the ponder move guess is right 90% of the time or more.
So if your new program is not at least 70 Elo stronger than the old, it is definitely better to use the old. Maybe as much as 100 Elo.

Let's do some math here...
Engine A is pondering with 100% right guesses.
Engine B is not pondering at all.
Two processes are used and 40 moves in 40 minues.
B is using 1 minute for each move as a mean value. Let's say exactly 1 minute/move.
A is pondering 1 minute and thinking 1 minute (without sofisticated time usage).
This gives exactly doubbled time and full usage of the time alotted for both. No doubt.
The question is what could change this?
One possiblity: Let's say that they are using smart time algorithms, thinking longer on hard positions.
Could this favour A? I'm not sure but maybe. I can't see how this is better than the first case described. Well, at least A have more time to play with. It's like how richer are getting even richer because of all the money available when needed.
Anyway, if A is really earning more than doubbled time just because of having more time to play with, then doubbling the n/s in your program would give you the same advantage just by having more nodes to play with per second.
I doubt it. Do you have som figures?
/Peter
PS. The 70 ELO earned by doubbling the time is an very old SSDF observation. I'm not sure it is even true anymore and at least not for all playing levels.

If you make your move and think on the opponent's move, it will be a doubling of time.
But what is usually done is speculation. You already make not only your move but also your opponent's expected response and that is what you ponder on. Then, if the opponent makes the move you guessed, you are way ahead. This is much better (for instance) than simply thinking twice as long and then moving or doubling the CPU speed.
Of course, if you guess the opponent move wrong all the time, yous should not do it that way. Probably, you have problems in you evaluation.
Every study I have seen still holds (approximately).

I don't follow you here.
We are comparing pondering with twice the time.
Let's stay with the 100% hit rate just to make it easier to compare. Lower hit rate is worse for pondering.
Case 1: The pondering case. Engine A has pondering and not B. 1 minute/move.
Case 2: Engine A gets 2 minutes/move and no pondering, B gets 1 minute/move without pondering.
In Case 1, A will guess the opponents move for 1 minute and if it's right he have another minute to use (or stop if he wants).
In Case 2, A will get the opponents move for free (no guessing) and will now have 2 minutes to use. Guessing is not even needed to know the opponent move.
You can't do any type of time speculation in Case 1 that you can't do in Case 2. On the contrary, in Case 1 you can't move until the opponent moves. In Case 2 you can move whenever you like.
/Peter

[Sune Fischer]

Geschrieben von:/Posted by: Sune Fischer at 06 April 2004 00:07:16:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Peter Fendrich at 05 April 2004 23:40:18:

Usually you can return instantly if you have already pondered longer than
what you intend to use on this move (unless if you're failing low or something).
I'm not sure the double time would be a good simulation of this.
The question is what happens if they start off the game by using the same amount of time, but one engine saves up a lot and has maybe 3-4 times as much on the clock for the endgame?
Is this equivalent to having twice as much time in the entire game?

I think it's best to stick to the 100% hit rate example - it's easier. That would be almost the same as having two minutes without pondering.
When B moves you have the option to continue your thinking or to stop. Exactly as if you had two minutes and no pondering.
Moving instantly is not for free. You have more time to use to get a better decision.
The 2 minute case as opposed to pondering is even slightly better because with a smart time algorithm you can sometimes move after 30 seconds if wanted. In the pondering alternative you can't move until the opponent moves.
Yes it is.
Saving time to later usage is taking a risk. It's not for free compared to use the time. Game theoretical it's even wrong to not use more of the time in the beginning of the game and save it for later usage. The game could end before that.
/Peter

If I had planned to think for 30 seconds and he lets me ponder for 50 there is no harm in that, more is simply better.
But by moving instantly you keep him on a time pressure.
If you think for 1 minute longer he in turn will be able to think on your clock, thus you're not just giving yourself more time, you're also giving your opponent more time.
Perhaps I missed your point?
How do you know?
This is not trivially true to me.
One could imagine for instance that a slight inaccuracy in the endgame
is generally more devastating than in the middlegame.
Also generally one tend to search deeper in the endgame, ie. you might get 2
more plies in some endgames versus half a ply in middle game.
I can't easily convince myself that all these factors should cancel out.
Yes that's correct, however using all your time on the first move isn't good
either, the truth is somewhere inbetween.
-S.

[Dann Corbit]

Geschrieben von:/Posted by: Dann Corbit at 06 April 2004 00:17:03:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Peter Fendrich at 06 April 2004 00:06:52:

------------------------------------
The Baron 1.3.0b3 1101 3.0/4
Movei 0.08.178 0010 1.0/4
------------------------------------
Phalanx XXII 1111 4.0/4 !!
Ufim 4.04 0000 0.0/4
------------------------------------
Terra 3.3b10 ½1½1 3.0/4
Frenzee 155 ½0½0 1.0/4
------------------------------------

Of course I am beaming!


Oh my goodness, Phalanx awakens. The world may never be the same.

Terra is proceeding according to plan...

No surprises here.

Frenzee is going to do really bad, it is playing with ponder OFF.
Leo, I'm starting to think it might be better to use the old v.146 with pondering.
I think when only one engine ponders the other engine will always get into a real time jam.
-S.

A 50% time penalty is 70 Elo (according to many estimates). There are some indications that pondering actually does better than a 50% horsepower increase and especially if the ponder move guess is right 90% of the time or more.
So if your new program is not at least 70 Elo stronger than the old, it is definitely better to use the old. Maybe as much as 100 Elo.

Let's do some math here...
Engine A is pondering with 100% right guesses.
Engine B is not pondering at all.
Two processes are used and 40 moves in 40 minues.
B is using 1 minute for each move as a mean value. Let's say exactly 1 minute/move.
A is pondering 1 minute and thinking 1 minute (without sofisticated time usage).
This gives exactly doubbled time and full usage of the time alotted for both. No doubt.
The question is what could change this?
One possiblity: Let's say that they are using smart time algorithms, thinking longer on hard positions.
Could this favour A? I'm not sure but maybe. I can't see how this is better than the first case described. Well, at least A have more time to play with. It's like how richer are getting even richer because of all the money available when needed.
Anyway, if A is really earning more than doubbled time just because of having more time to play with, then doubbling the n/s in your program would give you the same advantage just by having more nodes to play with per second.
I doubt it. Do you have som figures?
/Peter
PS. The 70 ELO earned by doubbling the time is an very old SSDF observation. I'm not sure it is even true anymore and at least not for all playing levels.

If you make your move and think on the opponent's move, it will be a doubling of time.
But what is usually done is speculation. You already make not only your move but also your opponent's expected response and that is what you ponder on. Then, if the opponent makes the move you guessed, you are way ahead. This is much better (for instance) than simply thinking twice as long and then moving or doubling the CPU speed.
Of course, if you guess the opponent move wrong all the time, yous should not do it that way. Probably, you have problems in you evaluation.
Every study I have seen still holds (approximately).

I don't follow you here.
We are comparing pondering with twice the time.
Let's stay with the 100% hit rate just to make it easier to compare. Lower hit rate is worse for pondering.
Case 1: The pondering case. Engine A has pondering and not B. 1 minute/move.
Case 2: Engine A gets 2 minutes/move and no pondering, B gets 1 minute/move without pondering.
In Case 1, A will guess the opponents move for 1 minute and if it's right he have another minute to use (or stop if he wants).
In Case 2, A will get the opponents move for free (no guessing) and will now have 2 minutes to use. Guessing is not even needed to know the opponent move.
You can't do any type of time speculation in Case 1 that you can't do in Case 2. On the contrary, in Case 1 you can't move until the opponent moves. In Case 2 you can move whenever you like.
/Peter

Two plies forward is a lot of search time, especially for long and deep searches.
Speculation is definitely a lot better if you get it right. That's why every engine that ponders does it that way. At least every one I examined.



my ftp site {remove http:// unless you like error messages}

[Sune Fischer]

Geschrieben von:/Posted by: Sune Fischer at 06 April 2004 00:24:52:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Peter Fendrich at 06 April 2004 00:06:52:

If you make your move and think on the opponent's move, it will be a doubling of time.
But what is usually done is speculation. You already make not only your move but also your opponent's expected response and that is what you ponder on. Then, if the opponent makes the move you guessed, you are way ahead. This is much better (for instance) than simply thinking twice as long and then moving or doubling the CPU speed.
Of course, if you guess the opponent move wrong all the time, yous should not do it that way. Probably, you have problems in you evaluation.

PS. The 70 ELO earned by doubbling the time is an very old SSDF observation. I'm not sure it is even true anymore and at least not for all playing levels.

Every study I have seen still holds (approximately).

I don't follow you here.
We are comparing pondering with twice the time.
Let's stay with the 100% hit rate just to make it easier to compare. Lower hit rate is worse for pondering.
Case 1: The pondering case. Engine A has pondering and not B. 1 minute/move.
Case 2: Engine A gets 2 minutes/move and no pondering, B gets 1 minute/move without pondering.
In Case 1, A will guess the opponents move for 1 minute and if it's right he have another minute to use (or stop if he wants).
In Case 2, A will get the opponents move for free (no guessing) and will now have 2 minutes to use. Guessing is not even needed to know the opponent move.
You can't do any type of time speculation in Case 1 that you can't do in Case 2. On the contrary, in Case 1 you can't move until the opponent moves. In Case 2 you can move whenever you like.
/Peter

Suppose we consider real play for a moment, how do you applyu case 2
in real play?
In real play you must assume that your opponent also ponders and will get about
as many ponder hits as you, so you cannot afford to just use twice the time.
In fact I don't really see a reason to have different time managers for ponder
and non-ponder modes. Things seem to average out by symmetry.
-S.

[Peter Fendrich]

Geschrieben von:/Posted by: Peter Fendrich at 06 April 2004 00:25:40:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Sune Fischer at 06 April 2004 00:07:16:

Usually you can return instantly if you have already pondered longer than
what you intend to use on this move (unless if you're failing low or something).
I'm not sure the double time would be a good simulation of this.
The question is what happens if they start off the game by using the same amount of time, but one engine saves up a lot and has maybe 3-4 times as much on the clock for the endgame?
Is this equivalent to having twice as much time in the entire game?

I think it's best to stick to the 100% hit rate example - it's easier. That would be almost the same as having two minutes without pondering.
When B moves you have the option to continue your thinking or to stop. Exactly as if you had two minutes and no pondering.
Moving instantly is not for free. You have more time to use to get a better decision.
The 2 minute case as opposed to pondering is even slightly better because with a smart time algorithm you can sometimes move after 30 seconds if wanted. In the pondering alternative you can't move until the opponent moves.
Yes it is.
Saving time to later usage is taking a risk. It's not for free compared to use the time. Game theoretical it's even wrong to not use more of the time in the beginning of the game and save it for later usage. The game could end before that.
/Peter

If I had planned to think for 30 seconds and he lets me ponder for 50 there is no harm in that, more is simply better.
But by moving instantly you keep him on a time pressure.
If you think for 1 minute longer he in turn will be able to think on your clock, thus you're not just giving yourself more time, you're also giving your opponent more time.
Perhaps I missed your point?
How do you know?
This is not trivially true to me.
One could imagine for instance that a slight inaccuracy in the endgame
is generally more devastating than in the middlegame.
Also generally one tend to search deeper in the endgame, ie. you might get 2
more plies in some endgames versus half a ply in middle game.
I can't easily convince myself that all these factors should cancel out.
Yes that's correct, however using all your time on the first move isn't good
either, the truth is somewhere inbetween.
-S.

If you have two minutes to use in whatever way you want it must be better compared to be forced to use some part of that time.
Absolutely not. B isn't pondering he is just waiting.
Well, I lost you here. I'm still comparing pondering with not pondering and trying to see how pondering could even be better than twice the time.
Disposing 2 minutes must be better than disposing 1 minute while guessing the opponents move and then disposing another minute.

[Sune Fischer]

Geschrieben von:/Posted by: Sune Fischer at 06 April 2004 00:33:49:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Peter Fendrich at 06 April 2004 00:25:40:

If I had planned to think for 30 seconds and he lets me ponder for 50 there is no harm in that, more is simply better.
But by moving instantly you keep him on a time pressure.
If you think for 1 minute longer he in turn will be able to think on your clock, thus you're not just giving yourself more time, you're also giving your opponent more time.
Perhaps I missed your point?

I'm not sure the double time would be a good simulation of this.
The question is what happens if they start off the game by using the same amount of time, but one engine saves up a lot and has maybe 3-4 times as much on the clock for the endgame?

If you have two minutes to use in whatever way you want it must be better compared to be forced to use some part of that time.
Absolutely not. B isn't pondering he is just waiting.
Well, I lost you here. I'm still comparing pondering with not pondering and trying to see how pondering could even be better than twice the time.
Disposing 2 minutes must be better than disposing 1 minute while guessing the opponents move and then disposing another minute.

Okay, I see.
As I just wrote in the other post I don't understand how you are going
to apply this "use double time" principle in a real game where your
opponent also ponders.
It's going to hit you twice as hard everything he gets a ponder hit, then
you must think for another 2 minutes and he will happily ponder away on your
time.
So you are forced to something strange like use twice time when I ponder hit
him and use half time when he ponder hits me.
This is not going to work well IMO.
-S.

[Peter Fendrich]

Geschrieben von:/Posted by: Peter Fendrich at 06 April 2004 00:38:44:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Dann Corbit at 06 April 2004 00:17:03:

------------------------------------
The Baron 1.3.0b3 1101 3.0/4
Movei 0.08.178 0010 1.0/4
------------------------------------
Phalanx XXII 1111 4.0/4 !!
Ufim 4.04 0000 0.0/4
------------------------------------
Terra 3.3b10 ½1½1 3.0/4
Frenzee 155 ½0½0 1.0/4
------------------------------------

Of course I am beaming!


Oh my goodness, Phalanx awakens. The world may never be the same.

Terra is proceeding according to plan...

No surprises here.

Frenzee is going to do really bad, it is playing with ponder OFF.
Leo, I'm starting to think it might be better to use the old v.146 with pondering.
I think when only one engine ponders the other engine will always get into a real time jam.
-S.

A 50% time penalty is 70 Elo (according to many estimates). There are some indications that pondering actually does better than a 50% horsepower increase and especially if the ponder move guess is right 90% of the time or more.
So if your new program is not at least 70 Elo stronger than the old, it is definitely better to use the old. Maybe as much as 100 Elo.

Let's do some math here...
Engine A is pondering with 100% right guesses.
Engine B is not pondering at all.
Two processes are used and 40 moves in 40 minues.
B is using 1 minute for each move as a mean value. Let's say exactly 1 minute/move.
A is pondering 1 minute and thinking 1 minute (without sofisticated time usage).
This gives exactly doubbled time and full usage of the time alotted for both. No doubt.
The question is what could change this?
One possiblity: Let's say that they are using smart time algorithms, thinking longer on hard positions.
Could this favour A? I'm not sure but maybe. I can't see how this is better than the first case described. Well, at least A have more time to play with. It's like how richer are getting even richer because of all the money available when needed.
Anyway, if A is really earning more than doubbled time just because of having more time to play with, then doubbling the n/s in your program would give you the same advantage just by having more nodes to play with per second.
I doubt it. Do you have som figures?
/Peter
PS. The 70 ELO earned by doubbling the time is an very old SSDF observation. I'm not sure it is even true anymore and at least not for all playing levels.

If you make your move and think on the opponent's move, it will be a doubling of time.
But what is usually done is speculation. You already make not only your move but also your opponent's expected response and that is what you ponder on. Then, if the opponent makes the move you guessed, you are way ahead. This is much better (for instance) than simply thinking twice as long and then moving or doubling the CPU speed.
Of course, if you guess the opponent move wrong all the time, yous should not do it that way. Probably, you have problems in you evaluation.
Every study I have seen still holds (approximately).

I don't follow you here.
We are comparing pondering with twice the time.
Let's stay with the 100% hit rate just to make it easier to compare. Lower hit rate is worse for pondering.
Case 1: The pondering case. Engine A has pondering and not B. 1 minute/move.
Case 2: Engine A gets 2 minutes/move and no pondering, B gets 1 minute/move without pondering.
In Case 1, A will guess the opponents move for 1 minute and if it's right he have another minute to use (or stop if he wants).
In Case 2, A will get the opponents move for free (no guessing) and will now have 2 minutes to use. Guessing is not even needed to know the opponent move.
You can't do any type of time speculation in Case 1 that you can't do in Case 2. On the contrary, in Case 1 you can't move until the opponent moves. In Case 2 you can move whenever you like.
/Peter

Two plies forward is a lot of search time, especially for long and deep searches.
Speculation is definitely a lot better if you get it right. That's why every engine that ponders does it that way. At least every one I examined.

You are missing something.
The claim is that pondering can even be better than having twice the time.
There are no favaurable speculations comparing pondering with having twice the time...
How can disposing 1 minute guessing the opponents move plus having another minute knowing the oppponents move be better than disposing 2 minutes knowing the opponents move from start?
One example:
Case 1:
I move 1.e4 from the start position and start pondering on opponents e5.
After 1 minute the opponent moves 1... e5.
Now I have another minute to dispose (if my speculation was right).
Case 2:
I move 1.e4 from the start position and waits for the opponent to move.
The opponent moves 1...e5 Now I have 2 minutes to dispose. No speculation necessary. I know the opponents move for sure.
How can case 1 be better than case 2 even with 100% hit rate?
/Peter

[Peter Fendrich]

Geschrieben von:/Posted by: Peter Fendrich at 06 April 2004 00:47:04:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Sune Fischer at 06 April 2004 00:24:52:

If you make your move and think on the opponent's move, it will be a doubling of time.
But what is usually done is speculation. You already make not only your move but also your opponent's expected response and that is what you ponder on. Then, if the opponent makes the move you guessed, you are way ahead. This is much better (for instance) than simply thinking twice as long and then moving or doubling the CPU speed.
Of course, if you guess the opponent move wrong all the time, yous should not do it that way. Probably, you have problems in you evaluation.

PS. The 70 ELO earned by doubbling the time is an very old SSDF observation. I'm not sure it is even true anymore and at least not for all playing levels.

Every study I have seen still holds (approximately).

I don't follow you here.
We are comparing pondering with twice the time.
Let's stay with the 100% hit rate just to make it easier to compare. Lower hit rate is worse for pondering.
Case 1: The pondering case. Engine A has pondering and not B. 1 minute/move.
Case 2: Engine A gets 2 minutes/move and no pondering, B gets 1 minute/move without pondering.
In Case 1, A will guess the opponents move for 1 minute and if it's right he have another minute to use (or stop if he wants).
In Case 2, A will get the opponents move for free (no guessing) and will now have 2 minutes to use. Guessing is not even needed to know the opponent move.
You can't do any type of time speculation in Case 1 that you can't do in Case 2. On the contrary, in Case 1 you can't move until the opponent moves. In Case 2 you can move whenever you like.
/Peter

Suppose we consider real play for a moment, how do you applyu case 2
in real play?
In real play you must assume that your opponent also ponders and will get about
as many ponder hits as you, so you cannot afford to just use twice the time.
In fact I don't really see a reason to have different time managers for ponder
and non-ponder modes. Things seem to average out by symmetry.
-S.

Yes, but the claim is that pondering can get as much as 100 ELO but twice the time only gets 70 ELO. That's why I comparing pondering with non-pondering.
It just can't be better with pondering. See my example in the end of:
http://f11.parsimony.net/forum16635/messages/64635.htm
...otherwise I'm very very sleepy...
/Peter

[Sune Fischer]

Geschrieben von:/Posted by: Sune Fischer at 06 April 2004 00:47:16:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Peter Fendrich at 06 April 2004 00:38:44:

------------------------------------
The Baron 1.3.0b3 1101 3.0/4
Movei 0.08.178 0010 1.0/4
------------------------------------
Phalanx XXII 1111 4.0/4 !!
Ufim 4.04 0000 0.0/4
------------------------------------
Terra 3.3b10 ½1½1 3.0/4
Frenzee 155 ½0½0 1.0/4
------------------------------------

Of course I am beaming!


Oh my goodness, Phalanx awakens. The world may never be the same.

Terra is proceeding according to plan...

No surprises here.

Frenzee is going to do really bad, it is playing with ponder OFF.
Leo, I'm starting to think it might be better to use the old v.146 with pondering.
I think when only one engine ponders the other engine will always get into a real time jam.
-S.

A 50% time penalty is 70 Elo (according to many estimates). There are some indications that pondering actually does better than a 50% horsepower increase and especially if the ponder move guess is right 90% of the time or more.
So if your new program is not at least 70 Elo stronger than the old, it is definitely better to use the old. Maybe as much as 100 Elo.

Let's do some math here...
Engine A is pondering with 100% right guesses.
Engine B is not pondering at all.
Two processes are used and 40 moves in 40 minues.
B is using 1 minute for each move as a mean value. Let's say exactly 1 minute/move.
A is pondering 1 minute and thinking 1 minute (without sofisticated time usage).
This gives exactly doubbled time and full usage of the time alotted for both. No doubt.
The question is what could change this?
One possiblity: Let's say that they are using smart time algorithms, thinking longer on hard positions.
Could this favour A? I'm not sure but maybe. I can't see how this is better than the first case described. Well, at least A have more time to play with. It's like how richer are getting even richer because of all the money available when needed.
Anyway, if A is really earning more than doubbled time just because of having more time to play with, then doubbling the n/s in your program would give you the same advantage just by having more nodes to play with per second.
I doubt it. Do you have som figures?
/Peter
PS. The 70 ELO earned by doubbling the time is an very old SSDF observation. I'm not sure it is even true anymore and at least not for all playing levels.

If you make your move and think on the opponent's move, it will be a doubling of time.
But what is usually done is speculation. You already make not only your move but also your opponent's expected response and that is what you ponder on. Then, if the opponent makes the move you guessed, you are way ahead. This is much better (for instance) than simply thinking twice as long and then moving or doubling the CPU speed.
Of course, if you guess the opponent move wrong all the time, yous should not do it that way. Probably, you have problems in you evaluation.
Every study I have seen still holds (approximately).

I don't follow you here.
We are comparing pondering with twice the time.
Let's stay with the 100% hit rate just to make it easier to compare. Lower hit rate is worse for pondering.
Case 1: The pondering case. Engine A has pondering and not B. 1 minute/move.
Case 2: Engine A gets 2 minutes/move and no pondering, B gets 1 minute/move without pondering.
In Case 1, A will guess the opponents move for 1 minute and if it's right he have another minute to use (or stop if he wants).
In Case 2, A will get the opponents move for free (no guessing) and will now have 2 minutes to use. Guessing is not even needed to know the opponent move.
You can't do any type of time speculation in Case 1 that you can't do in Case 2. On the contrary, in Case 1 you can't move until the opponent moves. In Case 2 you can move whenever you like.
/Peter

Two plies forward is a lot of search time, especially for long and deep searches.
Speculation is definitely a lot better if you get it right. That's why every engine that ponders does it that way. At least every one I examined.

You are missing something.
The claim is that pondering can even be better than having twice the time.
There are no favaurable speculations comparing pondering with having twice the time...
How can disposing 1 minute guessing the opponents move plus having another minute knowing the oppponents move be better than disposing 2 minutes knowing the opponents move from start?
One example:
Case 1:
I move 1.e4 from the start position and start pondering on opponents e5.
After 1 minute the opponent moves 1... e5.
Now I have another minute to dispose (if my speculation was right).
Case 2:
I move 1.e4 from the start position and waits for the opponent to move.
The opponent moves 1...e5 Now I have 2 minutes to dispose. No speculation necessary. I know the opponents move for sure.
How can case 1 be better than case 2 even with 100% hit rate?
/Peter

Ok, now I finally get
Yes I agree that double time must be an upper limit for what can be achieved
with pondering.
I think this can even be proven.
For one thing you wouldn't have to save up time before you could use it, you
could start using the extra time from move 1 if needed.
Ignore my other posts, I thought you were suggesting some kind of alternative
ponder strategy
-S.

[Uri Blass]

Geschrieben von:/Posted by: Uri Blass at 06 April 2004 08:11:24:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Roger Brown at 05 April 2004 19:49:19:

------------------------------------
The Baron 1.3.0b3 1101 3.0/4
Movei 0.08.178 0010 1.0/4
------------------------------------

Of course I am beaming!

This result suggest that maybe I did a mistake when I asked Leo to use 178 and not 178d.
178d got clearly better result against Ufim (3.5-0.5)
I decided to ask Leo to use 178 and not 178d not only based on my nunn match results when 178 beated 178d but also because of the following results thanks to Dan Honeycutt.
1: Movei00_8_178 11.5/20 ·········· 00=111=110 =1==1==001 106.75
2: Phalanx 10.0/20 11=000=001 ·········· =11011=001 97.00
3: Movei00_8_178d 8.5/20 =0==0==110 =00100=110 ·········· 91.75
30 games played / Tournament finished
Name of the tournament: Test04_02
Site/ Country: APPLE, United States
Level: Tournament 40/7
I did not care about result against previous version but about results against Phalanx when 00_8_178 beated it 6-4 when 00_8_178d lost 6-4
The problem is that the same tester sent me later the following results
1: Pepito 27.5/40 ·········· 011101==== ==11101111 1001=11=10 001111=111 497.00
2: Movei00_8_178d 19.5/40 100010==== ·········· 0=1=000011 1=110=0111 =111010000 384.00
3: Movei00_8_178 19.0/40 ==00010000 1=0=111100 ·········· 110101011= 1=10=0=001 361.00
4: Phalanx 18.0/40 0110=00=01 0=001=1000 001010100= ·········· 11=111==01 355.00
5: Movei00_8_174 16.0/40 110000=000 =000101111 0=01=1=110 00=000==10 ·········· 325.50
100 games played / Tournament finished
Name of the tournament: Test04_03
Site/ Country: APPLE, United States
Level: Tournament 40/5

You can see now that 00_8_178d lost 6-4 against pepito when 00_8_178 lost 8-2 and it also beated phalanx 6.5-3.5 when 00_8_178 beated it only 6-4 so if I do not count games against previous versions 00_8_178d is now 1/2 point better.
Uri

[Richard Pijl]

Geschrieben von:/Posted by: Richard Pijl at 06 April 2004 09:28:40:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Uri Blass at 06 April 2004 08:11:24:

------------------------------------
The Baron 1.3.0b3 1101 3.0/4
Movei 0.08.178 0010 1.0/4
------------------------------------

Of course I am beaming!

This result suggest that maybe I did a mistake when I asked Leo to use 178 and not 178d.

Perhaps the Baron did improve a little?
From my own tests, this version does perform significantly better than the Baron 1.2.1 (about 40 ELO points on 40/5 and ponder on with identical books and >300 games played against 80 different opponents) or the Baron 1.3.0b1.
If anything can be concluded from 4 games, it is of course that anything can happen ...
Before anyone asks: the release of this version will be after the tournament in Leiden at the end of this month .
Richard.

[Leo Dijksman]

Geschrieben von:/Posted by: Leo Dijksman at 06 April 2004 20:18:53:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Sune Fischer at 05 April 2004 23:44:36:

Version 149 has the same threaded design as 151 and 155, so it will have
the same bug.

Thats strange, I cannot remember I had problems with 149 here:

It's a random bug, perhaps they weren't very good at ponder guessing frenzee's
move?

Just to be sure I have start 146 and 149 playing 2 games (at 40/5, ponder=on) vs each other and 16 other opponents, I will send the result tomorrow.
The first one is Frenzee 149 vs 146 2-0 )
Best wishes,
Leo.

Okay Leo, if you can magicly get 149 to play without hanging, then please do!

-S.

Hi Sune,
The result:
Frenzee test
MP2600, 2004.04.05 - 2004.04.06
Score Fr Fr Am Go Ru Ar Am Ph Bu Uf Re Sc Ce Kn Ch Es Sl Te
---------------------------------------------------------------------------------------
1: Frenzee 149 19.0 / 34 XX 11 00 00 00 00 00 01 01 1= 10 =1 11 10 11 11 11 11
2: Frenzee 146 13.5 / 34 00 XX 00 00 =0 01 01 00 01 0= 1= 01 =0 11 10 10 01 =1
3: Amyan 1.592 4.0 / 4 11 11 XX .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
4: Gothmog 0.4.6 4.0 / 4 11 11 .. XX .. .. .. .. .. .. .. .. .. .. .. .. .. ..
5: Ruffian 0.7.6 3.5 / 4 11 =1 .. .. XX .. .. .. .. .. .. .. .. .. .. .. .. ..
6: Arasan 7.2 3.0 / 4 11 10 .. .. .. XX .. .. .. .. .. .. .. .. .. .. .. ..
7: Amateur 2.71 3.0 / 4 11 10 .. .. .. .. XX .. .. .. .. .. .. .. .. .. .. ..
8: Phalanx XXII 3.0 / 4 10 11 .. .. .. .. .. XX .. .. .. .. .. .. .. .. .. ..
9: Butcher 1.42c 2.0 / 4 10 10 .. .. .. .. .. .. XX .. .. .. .. .. .. .. .. ..
10: Ufim 4.04 2.0 / 4 0= 1= .. .. .. .. .. .. .. XX .. .. .. .. .. .. .. ..
11: Resp 0.19 1.5 / 4 01 0= .. .. .. .. .. .. .. .. XX .. .. .. .. .. .. ..
12: Scidlet 3.5a 1.5 / 4 =0 10 .. .. .. .. .. .. .. .. .. XX .. .. .. .. .. ..
13: Cerebro 1.23a 1.5 / 4 00 =1 .. .. .. .. .. .. .. .. .. .. XX .. .. .. .. ..
14: Knightx 1.81 1.0 / 4 01 00 .. .. .. .. .. .. .. .. .. .. .. XX .. .. .. ..
15: Chezzz 1.0.3 1.0 / 4 00 01 .. .. .. .. .. .. .. .. .. .. .. .. XX .. .. ..
16: Esc 1.16 1.0 / 4 00 01 .. .. .. .. .. .. .. .. .. .. .. .. .. XX .. ..
17: SlowChess 2.89a 1.0 / 4 00 10 .. .. .. .. .. .. .. .. .. .. .. .. .. .. XX ..
18: Terra 3.3b3 0.5 / 4 00 =0 .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. XX
---------------------------------------------------------------------------------------
66 games: +31 =7 -28
149 scored better but lost 2 games on time (in both cases no answer to a very fast pondermove of the opponent) but did not "hang".
Looking at the 2 timeloses I think I should use Frenzee 146 from now on?
Best wishes,
Leo.


WBEC Ridderkerk homepage.

[Sune Fischer]

Geschrieben von:/Posted by: Sune Fischer at 06 April 2004 21:39:49:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Leo Dijksman at 06 April 2004 20:18:53:

Looking at the 2 timeloses I think I should use Frenzee 146 from now on?
Best wishes,
Leo.

Thanks for running that little gauntlet, I think you must have too many computers
I think you probably should use the old stable version. There is not much
choice as the rules would soon disqualify the new version due to time loses.
-S.

[Leo Dijksman]

Geschrieben von:/Posted by: Leo Dijksman at 06 April 2004 22:58:11:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Sune Fischer at 06 April 2004 21:39:49:

Looking at the 2 timeloses I think I should use Frenzee 146 from now on?
Best wishes,
Leo.

Thanks for running that little gauntlet, I think you must have too many computers
I think you probably should use the old stable version. There is not much
choice as the rules would soon disqualify the new version due to time loses.
-S.

The alternative is one version with the strenght of 149 and as stable as 146 )
Best wishes,
Leo.



WBEC Ridderkerk homepage.

[Sune Fischer]

Geschrieben von:/Posted by: Sune Fischer at 07 April 2004 01:09:56:
Als Antwort auf:/In reply to: Re: WBEC Ridderkerk new results. geschrieben von:/posted by: Leo Dijksman at 06 April 2004 22:58:11:

Looking at the 2 timeloses I think I should use Frenzee 146 from now on?
Best wishes,
Leo.

Thanks for running that little gauntlet, I think you must have too many computers
I think you probably should use the old stable version. There is not much
choice as the rules would soon disqualify the new version due to time loses.
-S.

The alternative is one version with the strenght of 149 and as stable as 146 )
Best wishes,
Leo.

I have a dream Leo, a dream that stability and strength be united.
It will have to be a dream of the future, right now everything is in pieces
and it is scoring horribly.
My new floodfillers finally working great and then something else has broken,
arrrrgh! Isn't that just typical.
Oh well, what doesn't kill me makes me stronger, right.
-S.