Update to have Stockfish 17 Win-Draw-Loss Model

[MX6464]

Can anyone add the new Stockfish 17 Win-Draw-Loss model? https://github.com/official-stockfish/WDL_model

[niklasf]

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.

[robertnurnberg]

That is correct. What does/did python-chess use SF's wdl model for?

[MX6464]

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)))

[robertnurnberg]

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.