philihp / openskill.js / 25618940322

100%

push

github

Achieve bit-for-bit parity with openskill.py 6.x (#971)

* feat!: bit-for-bit Plackett-Luce parity with openskill.py 6.x

Two breaking changes align JS hyperparameter resolution and tau application
with the Python reference:

- src/constants.ts: defaults for sigma/beta/tau no longer derive from mu/z.
  They are now constants 25/3, 25/6, 25/300 — matching openskill.py's
  PlackettLuce(mu=25, sigma=25/3, beta=25/6, tau=25/300, kappa=1e-4).
- src/rate.ts: tau correction sqrt(sigma^2 + tau^2) now applies
  unconditionally with the resolved TAU (no-op when tau=0), mirroring
  Python's behaviour.

src/models/plackett-luce.ts also reorders the iDelta multiplication so
operation order matches Python (delta * sigma_sq/c**2 * gamma) and yields
bit-identical floats.

Existing test expectations regenerated against openskill==6.2.0; new
parity-plackett-luce.test.ts pins three scenarios (single doubles,
4-team free-for-all, 16-match tournament at mu=0/sigma=1) to Python's
exact float outputs via toStrictEqual.

Single-match rate() output is byte-identical to Python's
PlackettLuce.rate() for tested inputs. The 1024-match harness in
/tmp/openskill-compare/ shows residual deltas of <1e-12 absolute,
attributable to platform libm differences (V8 Math.exp vs CPython
math.exp) and not algorithmic divergence.

Bumps package.json to 5.0.0.

* feat!: bit-for-bit parity for all five models (CDF/PDF + epsilon)

Switch standard-normal CDF/PDF in src/statistics.ts from
@stdlib/stats-base-dists-normal-{cdf,pdf} to a direct erf-based
implementation matching CPython's statistics.NormalDist algorithm:
- phiMajor(x) = 0.5 * (1 + erf(x / sqrt(2)))
- phiMinor(x) = exp(x*x / -2) / sqrt(2*pi)

@stdlib/math-base-special-erf produces bit-identical floats to
math.erf for all tested inputs, so Thurstone-Mosteller (and the
predict-* helpers that call phiMajor) now bit-match Python.

Also fix the EPSILON/KAPPA conflation. Python's openskill.py uses
two separate constants:
- epsilon (default 0.... (continued)

93 of 93 branches covered (100.0%)

Branch coverage included in aggregate %.

21 of 21 new or added lines in 8 files covered. (100.0%)

292 of 292 relevant lines covered (100.0%)

122.52 hits per line