Activity

1. niklasf commented on Sep 27, 2024

   

   niklasf

   on Sep 27, 2024

   Owner

   More actions

   Looks like this time it's not just updating the parameters (official-stockfish/Stockfish@2054add), but also taking into account material counts: official-stockfish/Stockfish@9b92ada.
2. robertnurnberg commented on Sep 27, 2024

   

   robertnurnberg

   on Sep 27, 2024

   Contributor

   More actions

   That is correct. What does/did python-chess use SF's wdl model for?
3. MX6464 commented on Sep 27, 2024

   

   MX6464

   on Sep 27, 2024

   Last edited by MX6464

   Author

   More actions

   python-chess WDL calculations found here:

   python-chess/chess/engine.py

   Lines 446 to 588 in 5826ef5

   	def wdl(self, *, model: WdlModel = "sf", ply: int = 30) -> Wdl:
   	"""
   	Returns statistics for the expected outcome of this game, based on
   	a *model*, given that this score is reached at *ply*.
   	Scores have a total order, but it makes little sense to compute
   	the difference between two scores. For example, going from
   	``Cp(-100)`` to ``Cp(+100)`` is much more significant than going
   	from ``Cp(+300)`` to ``Cp(+500)``. It is better to compute differences
   	of the expectation values for the outcome of the game (based on winning
   	chances and drawing chances).
   	>>> Cp(100).wdl().expectation() - Cp(-100).wdl().expectation() # doctest: +ELLIPSIS
   	0.379...
   	>>> Cp(500).wdl().expectation() - Cp(300).wdl().expectation() # doctest: +ELLIPSIS
   	0.015...
   	:param model:
   	* ``sf``, the WDL model used by the latest Stockfish
   	(currently ``sf16``).
   	* ``sf16``, the WDL model used by Stockfish 16.
   	* ``sf15.1``, the WDL model used by Stockfish 15.1.
   	* ``sf15``, the WDL model used by Stockfish 15.
   	* ``sf14``, the WDL model used by Stockfish 14.
   	* ``sf12``, the WDL model used by Stockfish 12.
   	* ``lichess``, the win rate model used by Lichess.
   	Does not use *ply*, and does not consider drawing chances.
   	:param ply: The number of half-moves played since the starting
   	position. Models may scale scores slightly differently based on
   	this. Defaults to middle game.
   	"""
   	@abc.abstractmethod
   	def __neg__(self) -> Score:
   	...
   	@abc.abstractmethod
   	def __pos__(self) -> Score:
   	...
   	@abc.abstractmethod
   	def __abs__(self) -> Score:
   	...
   	def _score_tuple(self) -> Tuple[bool, bool, bool, int, Optional[int]]:
   	mate = self.mate()
   	return (
   	isinstance(self, MateGivenType),
   	mate is not None and mate > 0,
   	mate is None,
   	-(mate or 0),
   	self.score(),
   	)
   	def __eq__(self, other: object) -> bool:
   	if isinstance(other, Score):
   	return self._score_tuple() == other._score_tuple()
   	else:
   	return NotImplemented
   	def __lt__(self, other: object) -> bool:
   	if isinstance(other, Score):
   	return self._score_tuple() < other._score_tuple()
   	else:
   	return NotImplemented
   	def __le__(self, other: object) -> bool:
   	if isinstance(other, Score):
   	return self._score_tuple() <= other._score_tuple()
   	else:
   	return NotImplemented
   	def __gt__(self, other: object) -> bool:
   	if isinstance(other, Score):
   	return self._score_tuple() > other._score_tuple()
   	else:
   	return NotImplemented
   	def __ge__(self, other: object) -> bool:
   	if isinstance(other, Score):
   	return self._score_tuple() >= other._score_tuple()
   	else:
   	return NotImplemented
   	def _sf16_1_wins(cp: int, *, ply: int) -> int:
   	# https://github.com/official-stockfish/Stockfish/blob/sf_16.1/src/uci.cpp#L48
   	NormalizeToPawnValue = 356
   	# https://github.com/official-stockfish/Stockfish/blob/sf_16.1/src/uci.cpp#L383-L384
   	m = min(120, max(8, ply / 2 + 1)) / 32
   	a = (((-1.06249702 * m + 7.42016937) * m + 0.89425629) * m) + 348.60356174
   	b = (((-5.33122190 * m + 39.57831533) * m + -90.84473771) * m) + 123.40620748
   	x = min(4000, max(cp * NormalizeToPawnValue / 100, -4000))
   	return int(0.5 + 1000 / (1 + math.exp((a - x) / b)))
   	def _sf16_wins(cp: int, *, ply: int) -> int:
   	# https://github.com/official-stockfish/Stockfish/blob/sf_16/src/uci.h#L38
   	NormalizeToPawnValue = 328
   	# https://github.com/official-stockfish/Stockfish/blob/sf_16/src/uci.cpp#L200-L224
   	m = min(240, max(ply, 0)) / 64
   	a = (((0.38036525 * m + -2.82015070) * m + 23.17882135) * m) + 307.36768407
   	b = (((-2.29434733 * m + 13.27689788) * m + -14.26828904) * m) + 63.45318330
   	x = min(4000, max(cp * NormalizeToPawnValue / 100, -4000))
   	return int(0.5 + 1000 / (1 + math.exp((a - x) / b)))
   	def _sf15_1_wins(cp: int, *, ply: int) -> int:
   	# https://github.com/official-stockfish/Stockfish/blob/sf_15.1/src/uci.h#L38
   	NormalizeToPawnValue = 361
   	# https://github.com/official-stockfish/Stockfish/blob/sf_15.1/src/uci.cpp#L200-L224
   	m = min(240, max(ply, 0)) / 64
   	a = (((-0.58270499 * m + 2.68512549) * m + 15.24638015) * m) + 344.49745382
   	b = (((-2.65734562 * m + 15.96509799) * m + -20.69040836) * m) + 73.61029937
   	x = min(4000, max(cp * NormalizeToPawnValue / 100, -4000))
   	return int(0.5 + 1000 / (1 + math.exp((a - x) / b)))
   	def _sf15_wins(cp: int, *, ply: int) -> int:
   	# https://github.com/official-stockfish/Stockfish/blob/sf_15/src/uci.cpp#L200-L220
   	m = min(240, max(ply, 0)) / 64
   	a = (((-1.17202460e-1 * m + 5.94729104e-1) * m + 1.12065546e+1) * m) + 1.22606222e+2
   	b = (((-1.79066759 * m + 11.30759193) * m + -17.43677612) * m) + 36.47147479
   	x = min(2000, max(cp, -2000))
   	return int(0.5 + 1000 / (1 + math.exp((a - x) / b)))
   	def _sf14_wins(cp: int, *, ply: int) -> int:
   	# https://github.com/official-stockfish/Stockfish/blob/sf_14/src/uci.cpp#L200-L220
   	m = min(240, max(ply, 0)) / 64
   	a = (((-3.68389304 * m + 30.07065921) * m + -60.52878723) * m) + 149.53378557
   	b = (((-2.01818570 * m + 15.85685038) * m + -29.83452023) * m) + 47.59078827
   	x = min(2000, max(cp, -2000))
   	return int(0.5 + 1000 / (1 + math.exp((a - x) / b)))
   	def _sf12_wins(cp: int, *, ply: int) -> int:
   	# https://github.com/official-stockfish/Stockfish/blob/sf_12/src/uci.cpp#L198-L218
   	m = min(240, max(ply, 0)) / 64
   	a = (((-8.24404295 * m + 64.23892342) * m + -95.73056462) * m) + 153.86478679
   	b = (((-3.37154371 * m + 28.44489198) * m + -56.67657741) * m) + 72.05858751
   	x = min(1000, max(cp, -1000))
   	return int(0.5 + 1000 / (1 + math.exp((a - x) / b)))
   	def _lichess_raw_wins(cp: int) -> int:
   	# https://github.com/lichess-org/lila/pull/11148
   	# https://github.com/lichess-org/lila/blob/2242b0a08faa06e7be5508d338ede7bb09049777/modules/analyse/src/main/WinPercent.scala#L26-L30
   	return round(1000 / (1 + math.exp(-0.00368208 * cp)))

   👍React with 👍1Reacted by Robert Nürnberg
4. robertnurnberg commented on Sep 27, 2024

   

   robertnurnberg

   on Sep 27, 2024

   Last edited by robertnurnberg

   Contributor

   More actions

   Ok, so this needs to be based on material count now, so quite an invasive change. Also NormalizeToPawnValue needs to be replaced by a now.

   Edit: And the capping to [-4000,4000] can go.
5. niklasf

   added this to the v2.0.0 milestone on Oct 4, 2024