lnd.xprv/routing/probability_estimator.go
Joost Jager 1fac41deed
routing+routerrpc: improve prob. estimation for untried connections
This commit changes mission control to partially base the estimated
probability for untried connections on historical results obtained in
previous payment attempts. This incentivizes routing nodes to keep all
of their channels in good shape.
2019-10-22 15:52:38 +02:00

156 lines
6.0 KiB
Go

package routing
import (
"math"
"time"
"github.com/lightningnetwork/lnd/lnwire"
"github.com/lightningnetwork/lnd/routing/route"
)
// probabilityEstimator returns node and pair probabilities based on historical
// payment results.
type probabilityEstimator struct {
// penaltyHalfLife defines after how much time a penalized node or
// channel is back at 50% probability.
penaltyHalfLife time.Duration
// aprioriHopProbability is the assumed success probability of a hop in
// a route when no other information is available.
aprioriHopProbability float64
// aprioriWeight is a value in the range [0, 1] that defines to what
// extent historical results should be extrapolated to untried
// connections. Setting it to one will completely ignore historical
// results and always assume the configured a priori probability for
// untried connections. A value of zero will ignore the a priori
// probability completely and only base the probability on historical
// results, unless there are none available.
aprioriWeight float64
// prevSuccessProbability is the assumed probability for node pairs that
// successfully relayed the previous attempt.
prevSuccessProbability float64
}
// getNodeProbability calculates the probability for connections from a node
// that have not been tried before. The results parameter is a list of last
// payment results for that node.
func (p *probabilityEstimator) getNodeProbability(now time.Time,
results NodeResults, amt lnwire.MilliSatoshi) float64 {
// If the channel history is not to be taken into account, we can return
// early here with the configured a priori probability.
if p.aprioriWeight == 1 {
return p.aprioriHopProbability
}
// If there is no channel history, our best estimate is still the a
// priori probability.
if len(results) == 0 {
return p.aprioriHopProbability
}
// The value of the apriori weight is in the range [0, 1]. Convert it to
// a factor that properly expresses the intention of the weight in the
// following weight average calculation. When the apriori weight is 0,
// the apriori factor is also 0. This means it won't have any effect on
// the weighted average calculation below. When the apriori weight
// approaches 1, the apriori factor goes to infinity. It will heavily
// outweigh any observations that have been collected.
aprioriFactor := 1/(1-p.aprioriWeight) - 1
// Calculate a weighted average consisting of the apriori probability
// and historical observations. This is the part that incentivizes nodes
// to make sure that all (not just some) of their channels are in good
// shape. Senders will steer around nodes that have shown a few
// failures, even though there may be many channels still untried.
//
// If there is just a single observation and the apriori weight is 0,
// this single observation will totally determine the node probability.
// The node probability is returned for all other channels of the node.
// This means that one failure will lead to the success probability
// estimates for all other channels being 0 too. The probability for the
// channel that was tried will not even recover, because it is
// recovering to the node probability (which is zero). So one failure
// effectively prunes all channels of the node forever. This is the most
// aggressive way in which we can penalize nodes and unlikely to yield
// good results in a real network.
probabilitiesTotal := p.aprioriHopProbability * aprioriFactor
totalWeight := aprioriFactor
for _, result := range results {
age := now.Sub(result.timestamp)
switch {
// Weigh success with a constant high weight of 1. There is no
// decay.
case result.success:
totalWeight++
probabilitiesTotal += p.prevSuccessProbability
// Weigh failures in accordance with their age. The base
// probability of a failure is considered zero, so nothing needs
// to be added to probabilitiesTotal.
case amt >= result.minPenalizeAmt:
totalWeight += p.getWeight(age)
}
}
return probabilitiesTotal / totalWeight
}
// getWeight calculates a weight in the range [0, 1] that should be assigned to
// a payment result. Weight follows an exponential curve that starts at 1 when
// the result is fresh and asymptotically approaches zero over time. The rate at
// which this happens is controlled by the penaltyHalfLife parameter.
func (p *probabilityEstimator) getWeight(age time.Duration) float64 {
exp := -age.Hours() / p.penaltyHalfLife.Hours()
return math.Pow(2, exp)
}
// getPairProbability estimates the probability of successfully traversing to
// toNode based on historical payment outcomes for the from node. Those outcomes
// are passed in via the results parameter.
func (p *probabilityEstimator) getPairProbability(
now time.Time, results NodeResults,
toNode route.Vertex, amt lnwire.MilliSatoshi) float64 {
// Retrieve the last pair outcome.
lastPairResult, ok := results[toNode]
// If there is no history for this pair, return the node probability
// that is a probability estimate for untried channel.
if !ok {
return p.getNodeProbability(now, results, amt)
}
// For successes, we have a fixed (high) probability. Those pairs
// will be assumed good until proven otherwise.
if lastPairResult.success {
return p.prevSuccessProbability
}
nodeProbability := p.getNodeProbability(now, results, amt)
// Take into account a minimum penalize amount. For balance errors, a
// failure may be reported with such a minimum to prevent too aggressive
// penalization. If the current amount is smaller than the amount that
// previously triggered a failure, we act as if this is an untried
// channel.
if amt < lastPairResult.minPenalizeAmt {
return nodeProbability
}
timeSinceLastFailure := now.Sub(lastPairResult.timestamp)
// Calculate success probability based on the weight of the last
// failure. When the failure is fresh, its weight is 1 and we'll return
// probability 0. Over time the probability recovers to the node
// probability. It would be as if this channel was never tried before.
weight := p.getWeight(timeSinceLastFailure)
probability := nodeProbability * (1 - weight)
return probability
}