logaritmisk
e4ff46f45c
EP message cancellation can leave a Gaussian's precision (pi) a tiny negative value — round-off of exactly zero. mu()/sigma() only special-cased pi == 0, so sigma() computed 1/sqrt(pi) = NaN for pi < 0. That NaN flowed through the moment-space Sub in the game diff-chain and poisoned every skill in the slice once it grew past ~75 competitors, making converge() return all-NaN on real-scale histories (regression vs 0.1.0, which stored sigma directly). Guard pi <= 0.0 in both accessors (improper Gaussian: mu 0, sigma infinite), matching the existing pi == 0 handling. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
72 lines
2.3 KiB
Rust
72 lines
2.3 KiB
Rust
//! Regression: a single time slice with many distinct competitors must converge to finite
|
|
//! skills. Before the `pi <= 0` guard in `Gaussian::mu()/sigma()`, EP message cancellation
|
|
//! produced a tiny-negative precision whose `sigma() = 1/sqrt(pi)` was NaN, which the
|
|
//! moment-space `Sub` in the game chain propagated into every skill once the slice grew past
|
|
//! ~75 competitors (e.g. a real ranking dataset with hundreds of players).
|
|
use trueskill_tt::{ConstantDrift, ConvergenceOptions, EPSILON, History, ITERATIONS, NullObserver};
|
|
|
|
/// Tiny deterministic LCG — avoids a dev-dependency on `rand`.
|
|
struct Lcg(u64);
|
|
impl Lcg {
|
|
fn next(&mut self) -> u64 {
|
|
self.0 = self
|
|
.0
|
|
.wrapping_mul(6364136223846793005)
|
|
.wrapping_add(1442695040888963407);
|
|
self.0
|
|
}
|
|
fn below(&mut self, n: usize) -> usize {
|
|
(self.next() >> 33) as usize % n
|
|
}
|
|
fn coin(&mut self) -> bool {
|
|
self.next() & 1 == 0
|
|
}
|
|
}
|
|
|
|
fn nan_after_fit(players: usize) -> usize {
|
|
let mut h: History<i64, ConstantDrift, NullObserver, String> = History::builder_with_key()
|
|
.beta(1.0)
|
|
.sigma(6.0)
|
|
.drift(ConstantDrift(0.1))
|
|
.convergence(ConvergenceOptions {
|
|
max_iter: ITERATIONS,
|
|
epsilon: EPSILON,
|
|
..Default::default()
|
|
})
|
|
.build();
|
|
|
|
let ids: Vec<String> = (0..players).map(|i| format!("p{i:04}")).collect();
|
|
let mut rng = Lcg(1);
|
|
for _ in 0..(players * 4) {
|
|
let a = rng.below(players);
|
|
let mut b = rng.below(players - 1);
|
|
if b >= a {
|
|
b += 1;
|
|
}
|
|
let (w, l) = if rng.coin() { (a, b) } else { (b, a) };
|
|
h.record_winner(&ids[w], &ids[l], 0).unwrap();
|
|
}
|
|
h.converge().unwrap();
|
|
|
|
ids.iter()
|
|
.filter(|id| {
|
|
h.current_skill(id.as_str())
|
|
.map(|g| !g.mu().is_finite() || !g.sigma().is_finite())
|
|
.unwrap_or(true)
|
|
})
|
|
.count()
|
|
}
|
|
|
|
#[test]
|
|
fn many_competitors_converge_to_finite_skills() {
|
|
// The NaN regression onset was between 70 and 80 competitors; 250 is comfortably past it
|
|
// and in the range of a real ranking dataset.
|
|
for players in [12usize, 75, 150, 250] {
|
|
assert_eq!(
|
|
nan_after_fit(players),
|
|
0,
|
|
"{players}-competitor history produced NaN skills"
|
|
);
|
|
}
|
|
}
|