RepoMirrors
/
osu
mirror of https://github.com/ppy/osu
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Compute the upper bound on deviation with a 99% confidence interval

This commit is contained in:
nathen 2024-03-09 23:10:53 -05:00
parent 8a26cdaaab
commit caba0510db
1 changed files with 40 additions and 41 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

81
osu.Game.Rulesets.Taiko/Difficulty/TaikoPerformanceCalculator.cs
View File

@ -36,7 +36,7 @@ protected override PerformanceAttributes CreatePerformanceAttributes(ScoreInfo s
countOk = score.Statistics.GetValueOrDefault(HitResult.Ok);
countMeh = score.Statistics.GetValueOrDefault(HitResult.Meh);
countMiss = score.Statistics.GetValueOrDefault(HitResult.Miss);
estimatedUr = computeEstimatedUr(score, taikoAttributes);
estimatedUr = computeDeviationUpperBound(taikoAttributes) * 10;
// The effectiveMissCount is calculated by gaining a ratio for totalSuccessfulHits and increasing the miss penalty for shorter object counts lower than 1000.
if (totalSuccessfulHits > 0)
@ -103,7 +103,7 @@ private double computeAccuracyValue(ScoreInfo score, TaikoDifficultyAttributes a
if (attributes.GreatHitWindow <= 0 || estimatedUr == null)
return 0;
double accuracyValue = Math.Pow(65 / estimatedUr.Value, 1.1) * Math.Pow(attributes.StarRating, 0.4) * 100.0;
double accuracyValue = Math.Pow(70 / estimatedUr.Value, 1.1) * Math.Pow(attributes.StarRating, 0.4) * 100.0;
double lengthBonus = Math.Min(1.15, Math.Pow(totalHits / 1500.0, 0.3));
@ -115,10 +115,11 @@ private double computeAccuracyValue(ScoreInfo score, TaikoDifficultyAttributes a
}
/// <summary>
/// Calculates the tap deviation for a player using the OD, object count, and scores of 300s, 100s, and misses, with an assumed mean hit error of 0.
/// Consistency is ensured as identical SS scores on the same map and settings yield the same deviation.
/// Computes an upper bound on the player's tap deviation based on the OD, number of circles and sliders,
/// and the hit judgements, assuming the player's mean hit error is 0. The estimation is consistent in that
/// two SS scores on the same map with the same settings will always return the same deviation.
/// </summary>
private double? computeEstimatedUr(ScoreInfo score, TaikoDifficultyAttributes attributes)
private double? computeDeviationUpperBound(TaikoDifficultyAttributes attributes)
{
if (totalSuccessfulHits == 0 || attributes.GreatHitWindow <= 0)
return null;
@ -126,53 +127,51 @@ private double computeAccuracyValue(ScoreInfo score, TaikoDifficultyAttributes a
double h300 = attributes.GreatHitWindow;
double h100 = attributes.OkHitWindow;
// Determines the probability of a deviation leading to the score's hit evaluations. The curve's apex represents the most probable deviation.
double likelihoodGradient(double d)
const double z = 2.32634787404; // 99% critical value for the normal distribution (one-tailed).
// The upper bound on deviation, calculated with the ratio of 300s to 100s, and the great hit window.
double? calcDeviationGreatWindow()
{
if (d <= 0)
return 0;
if (countGreat == 0) return null;
double p300 = logDiff(0, logPcHit(h300, d));
double p100 = logDiff(logPcHit(h300, d), logPcHit(h100, d));
double p0 = logPcHit(h100, d);
double n = totalSuccessfulHits;
double gradient = Math.Exp(
(countGreat * p300
+ (countOk + 0.5) * p100
+ countMiss * p0) / totalHits
);
// Proportion of greats hit, ignoring misses.
double p = countGreat / n;
return -gradient;
// We can be 99% confident that p is at least this value.
double pLowerBound = (n * p + z * z / 2) / (n + z * z) - z / (n + z * z) * Math.Sqrt(n * p * (1 - p) + z * z / 4);
// We can be 99% confident that the deviation is not higher than:
return h300 / (Math.Sqrt(2) * SpecialFunctions.ErfInv(pLowerBound));
}
double deviation = FindMinimum.OfScalarFunction(likelihoodGradient, 30);
// The upper bound on deviation, calculated with the ratio of 300s + 100s to misses, and the good hit window.
// This will return a lower value than the first method when the number of 100s is high, but the miss count is low.
double? calcDeviationGoodWindow()
{
if (totalSuccessfulHits == 0) return null;
return deviation * 10;
double n = totalHits;
// Proportion of greats + goods hit.
double p = totalSuccessfulHits / n;
// We can be 99% confident that p is at least this value.
double pLowerBound = (n * p + z * z / 2) / (n + z * z) - z / (n + z * z) * Math.Sqrt(n * p * (1 - p) + z * z / 4);
// We can be 99% confident that the deviation is not higher than:
return h100 / (Math.Sqrt(2) * SpecialFunctions.ErfInv(pLowerBound));
}
if (calcDeviationGreatWindow() is null)
return calcDeviationGoodWindow();
return Math.Min(calcDeviationGreatWindow()!.Value, calcDeviationGoodWindow()!.Value);
}
private int totalHits => countGreat + countOk + countMeh + countMiss;
private int totalSuccessfulHits => countGreat + countOk + countMeh;
private double logPcHit(double x, double deviation) => logErfcApprox(x / (deviation * Math.Sqrt(2)));
// Utilises a numerical approximation to extend the computation range of ln(erfc(x)).
private double logErfcApprox(double x) => x <= 5
? Math.Log(SpecialFunctions.Erfc(x))
: -Math.Pow(x, 2) - Math.Log(x * Math.Sqrt(Math.PI)); // https://www.desmos.com/calculator/kdbxwxgf01
// Subtracts the base value of two logs, circumventing log rules that typically complicate subtraction of non-logarithmic values.
private double logDiff(double firstLog, double secondLog)
{
double maxVal = Math.Max(firstLog, secondLog);
// To avoid a NaN result, a check is performed to prevent subtraction of two negative infinity values.
if (double.IsNegativeInfinity(maxVal))
{
return maxVal;
}
return firstLog + SpecialFunctions.Log1p(-Math.Exp(-(firstLog - secondLog)));
}
}
}