737d7c7dfc
The ApplyUpdates() method on the reputation.DB interface acts like the similar Update() method, but can allow for applying the changes from multiple audit events, instead of only one. This will be necessary for the reputation write cache, which will batch up changes to each node's reputation in order to flush them periodically. Refs: https://github.com/storj/storj/issues/4601 Change-Id: I44cc47767ea2d9423166bb8fed080c8a11182041
70 lines
2.4 KiB
Go
70 lines
2.4 KiB
Go
// Copyright (C) 2022 Storj Labs, Inc.
|
|
// See LICENSE for copying information.
|
|
|
|
package reputation
|
|
|
|
import "math"
|
|
|
|
// UpdateReputation uses the Beta distribution model to determine a node's reputation.
|
|
// lambda is the "forgetting factor" which determines how much past info is kept when determining current reputation score.
|
|
// w is the normalization weight that affects how severely new updates affect the current reputation distribution.
|
|
func UpdateReputation(isSuccess bool, alpha, beta, lambda, w float64) (newAlpha, newBeta float64) {
|
|
// v is a single feedback value that allows us to update both alpha and beta
|
|
var v float64 = -1
|
|
if isSuccess {
|
|
v = 1
|
|
}
|
|
newAlpha = lambda*alpha + w*(1+v)/2
|
|
newBeta = lambda*beta + w*(1-v)/2
|
|
return newAlpha, newBeta
|
|
}
|
|
|
|
// UpdateReputationMultiple works like UpdateReputation, but applies multiple
|
|
// successive counts of an event type to the alpha and beta measures.
|
|
//
|
|
// With the arguments as named, applies 'count' successful audits. To apply negative
|
|
// audits, swap the alpha and beta parameters and return values.
|
|
//
|
|
//
|
|
// WARNING: GREEK LETTER MATH AHEAD
|
|
//
|
|
// Applying n successful audit results to an initial alpha value of α₀ gives a
|
|
// new α₁ value of:
|
|
//
|
|
// α₁ = λⁿα₀ + λⁿ⁻¹w + λⁿ⁻²w + ... + λ²w + λw + w
|
|
//
|
|
// The terms with w are the first n terms of a geometric series with coefficient
|
|
// w and common ratio λ. The closed form formula for the sum of those first n
|
|
// terms is (w(1-λⁿ) / (1-λ))
|
|
// (https://en.wikipedia.org/wiki/Geometric_series#Closed-form_formula).
|
|
// Adding the initial λⁿα₀ term, we get
|
|
//
|
|
// α₁ = λⁿα₀ + w(1-λⁿ) / (1-λ)
|
|
//
|
|
// The formula has the same structure for beta for n _failures_.
|
|
//
|
|
// β₁ = λⁿβ₀ + w(1-λⁿ) / (1-λ)
|
|
//
|
|
// For n _failures_,
|
|
//
|
|
// α₁ = λⁿα₀
|
|
//
|
|
// For n _successes_,
|
|
//
|
|
// β₁ = λⁿβ₀
|
|
//
|
|
func UpdateReputationMultiple(count int, alpha, beta, lambda, w float64) (newAlpha, newBeta float64) {
|
|
if lambda == 1 {
|
|
// special case: when the coefficient is 1, the closed-form formula is invalid
|
|
// (gives NaN because of a division by zero). Fortunately, the replacement
|
|
// formula in this case is even simpler.
|
|
newAlpha = alpha + w*float64(count)
|
|
newBeta = beta
|
|
} else {
|
|
lambdaPowN := math.Pow(lambda, float64(count))
|
|
newAlpha = lambdaPowN*alpha + w*(1-lambdaPowN)/(1-lambda)
|
|
newBeta = lambdaPowN * beta
|
|
}
|
|
return newAlpha, newBeta
|
|
}
|