PolyGlot and pondering

[Fabien Letouzey]

Geschrieben von:/Posted by: Fabien Letouzey at 13 April 2004 12:43:43:


Hello,
Let me mention again that Pondering does not work at all in PolyGlot.
There is a big incompatibility between xboard and UCI that I
overlooked. Because of that, I can't quickly solve the problem "the
right way".
I do understand, however, that pondering is very important for
tournaments.
Here I just want to discuss a possible work around that could be
available soon:
Because it is a timing (clock information) problem, I could let the
engine ponder normally then interrupt it even if the expected move is
played. A normal search would then be launched, with correct clock
information (UCI does not allow clock information to be updated when
the opponent plays the expected move).
Of course this is not real pondering!!! The search will have to start
from scratch, only hash-table information will be reused. I just want
to get your opinion about it (engine authors, tournament directors,
...). Is such a "crippled pondering" worth the development (quite
easy, it must be said)?
There is the risk that results are misinterpreted (engines would not
play at full strength). On the other hand, it's better that not
pondering at all ...
Fabien.

[Matthias Gemuh]

Geschrieben von:/Posted by: Matthias Gemuh at 13 April 2004 17:09:43:
Als Antwort auf:/In reply to: PolyGlot and pondering geschrieben von:/posted by: Fabien Letouzey at 13 April 2004 12:43:43:

Hello,
Let me mention again that Pondering does not work at all in PolyGlot.
There is a big incompatibility between xboard and UCI that I
overlooked. Because of that, I can't quickly solve the problem "the
right way".
I do understand, however, that pondering is very important for
tournaments.
Here I just want to discuss a possible work around that could be
available soon:
Because it is a timing (clock information) problem, I could let the
engine ponder normally then interrupt it even if the expected move is
played. A normal search would then be launched, with correct clock
information (UCI does not allow clock information to be updated when
the opponent plays the expected move).
Of course this is not real pondering!!! The search will have to start
from scratch, only hash-table information will be reused. I just want
to get your opinion about it (engine authors, tournament directors,
...). Is such a "crippled pondering" worth the development (quite
easy, it must be said)?
There is the risk that results are misinterpreted (engines would not
play at full strength). On the other hand, it's better that not
pondering at all ...
Fabien.

Perfect !! Implement that, please .
/Matthias.



BigLion + Taktix

[Tord Romstad]

Geschrieben von:/Posted by: Tord Romstad at 13 April 2004 17:14:31:
Als Antwort auf:/In reply to: PolyGlot and pondering geschrieben von:/posted by: Fabien Letouzey at 13 April 2004 12:43:43:

Hello,
Let me mention again that Pondering does not work at all in PolyGlot.
There is a big incompatibility between xboard and UCI that I
overlooked. Because of that, I can't quickly solve the problem "the
right way".
I do understand, however, that pondering is very important for
tournaments.
Here I just want to discuss a possible work around that could be
available soon:
Because it is a timing (clock information) problem, I could let the
engine ponder normally then interrupt it even if the expected move is
played. A normal search would then be launched, with correct clock
information (UCI does not allow clock information to be updated when
the opponent plays the expected move).
Of course this is not real pondering!!! The search will have to start
from scratch, only hash-table information will be reused. I just want
to get your opinion about it (engine authors, tournament directors,
...). Is such a "crippled pondering" worth the development (quite
easy, it must be said)?
There is the risk that results are misinterpreted (engines would not
play at full strength). On the other hand, it's better that not
pondering at all ...

Gothmog wouldn't be very happy. Whenever a new search is started, it increments
a generation counter. This has the consequence that all the information stored
in the hash table is used for move ordering and nothing else. In other words,
the only benefit my engine would have of your pseudo-pondering is slightly
improved move ordering.
Tord

[Fabien Letouzey]

Geschrieben von:/Posted by: Fabien Letouzey at 13 April 2004 17:26:37:
Als Antwort auf:/In reply to: Re: PolyGlot and pondering geschrieben von:/posted by: Tord Romstad at 13 April 2004 17:14:31:

Gothmog wouldn't be very happy. Whenever a new search is started, it increments
a generation counter. This has the consequence that all the information stored
in the hash table is used for move ordering and nothing else. In other words,
the only benefit my engine would have of your pseudo-pondering is slightly
improved move ordering.

Yes, but how does that compare to "ponder off"?
Fabien.

[Tord Romstad]

Geschrieben von:/Posted by: Tord Romstad at 13 April 2004 17:48:38:
Als Antwort auf:/In reply to: Re: PolyGlot and pondering geschrieben von:/posted by: Fabien Letouzey at 13 April 2004 17:26:37:

Gothmog wouldn't be very happy. Whenever a new search is started, it increments
a generation counter. This has the consequence that all the information stored
in the hash table is used for move ordering and nothing else. In other words,
the only benefit my engine would have of your pseudo-pondering is slightly
improved move ordering.

Yes, but how does that compare to "ponder off"?

Hard to say. When the pondering move is correct, it would improve the move
ordering slightly. When the pondering move is incorrect, the opposite would
happen (because information from the previous search has been overwritten
with less relevant information). I guess the difference compared to "ponder off"
would be very small.
Tord

[Fabien Letouzey]

Geschrieben von:/Posted by: Fabien Letouzey at 13 April 2004 17:55:11:
Als Antwort auf:/In reply to: Re: PolyGlot and pondering geschrieben von:/posted by: Tord Romstad at 13 April 2004 17:48:38:

Gothmog wouldn't be very happy. Whenever a new search is started, it increments
a generation counter. This has the consequence that all the information stored
in the hash table is used for move ordering and nothing else. In other words,
the only benefit my engine would have of your pseudo-pondering is slightly
improved move ordering.

Yes, but how does that compare to "ponder off"?

Hard to say. When the pondering move is correct, it would improve the move
ordering slightly. When the pondering move is incorrect, the opposite would
happen (because information from the previous search has been overwritten
with less relevant information). I guess the difference compared to "ponder off"
would be very small.

I see.
Then engine authors will have to be consulted for ponder-on tournaments. First to know if they find participating worthwhile at all, and then to choose between crippled pondering or no pondering at all!
Only bad choices, of course ...
Fabien.

[Peter Schäfer]

Geschrieben von:/Posted by: Peter Schäfer at 13 April 2004 19:27:37:
Als Antwort auf:/In reply to: Re: PolyGlot and pondering geschrieben von:/posted by: Fabien Letouzey at 13 April 2004 13:53:24:

Hello,
Let me mention again that Pondering does not work at all in PolyGlot.
There is a big incompatibility between xboard and UCI that I
overlooked. Because of that, I can't quickly solve the problem "the
right way".

what exactly is your problem ?

Remember that PolyGlot connects an xboard GUI with a UCI engine.
---
1) UCI does not allow clock information to be updated when the expected move is played by the opponent ("ponderhit" allows no argument).
=> correct data must be sent when pondering is started ("go ponder ...").
2) XBoard only sends clock info when the opponent plays (just before the engine turn).
IMO it should send data after each half move!
---
How can I accomodate with that?
The only solution I can think of is to handle my own clocks.
This would duplicate XBoard's work and would not be 100% accurate anyway.

I see...
My opinion: less accurate pondering is better than no pondering at all.
-- Peter

[Fabien Letouzey]

Geschrieben von:/Posted by: Fabien Letouzey at 14 April 2004 16:28:03:
Als Antwort auf:/In reply to: PolyGlot and pondering geschrieben von:/posted by: Fabien Letouzey at 13 April 2004 12:43:43:

Here I just want to discuss a possible work around that could be
available soon:
Because it is a timing (clock information) problem, I could let the
engine ponder normally then interrupt it even if the expected move is
played. A normal search would then be launched, with correct clock
information (UCI does not allow clock information to be updated when
the opponent plays the expected move).
Of course this is not real pondering!!! The search will have to start
from scratch, only hash-table information will be reused.

OK,
A new version with crippled pondering should be available tonight (european time) when Dann has time. It's not great of course, but hopefully will avoid the time-out debacle.
I have also fixed "Move Now" which was not working at all!
Fabien.

[Fabien Letouzey]

Geschrieben von:/Posted by: Fabien Letouzey at 14 April 2004 17:45:04:
Als Antwort auf:/In reply to: PolyGlot and pondering geschrieben von:/posted by: Fabien Letouzey at 13 April 2004 12:43:43:


Just to make sure UCI2WB suffers from the same problem.
Anybody can post (part of) a UCI2WB log file with ponder on?
Just 2 or 3 moves (no book is needed), ponder on, and also an expected move needs to be played.
Thanks,
Fabien.