* The maximal reduction in the probability is limited to 0.5 (previously
~0.05), such that we don't get too low apriori probabilities.
Otherwise, this may lead to a too strong selection of large (and maybe
expensive) channels. A two-hop path would get total probability
penalties of:
- 1000PPM/(0.6*0.6) = 2778 PPM in the unsaturated case
- 1000PPM/(0.6*(0.6*0.5)) = 5556 PPM in the saturated case, where the
second hop is saturated
The difference in PPM of 2778 PPM should be enough to bias towards the
first path.
* The smearing factor is reduced. Previously we had to keep a higher
smearing factor in order to make the capacity factor not go to zero
for high amounts, to still give a fully saturated channel a chance.
This is not needed anymore due to the capping to 0.5. A lower value of
the smearing factor lets us more precisely choose a capacity fraction
and the capacity factor is more neutral when it comes to intermediate
amounts.
We set a conservative default value for the capacity fraction, which
still has the effect of discarding exhausted channels, giving a
noticeable effect when about 90% of the capacity is being used.
We make the capacity factor configurable via an lnd.conf routerrpc
apriori parameter. The capacity factor trades off increased success
probability with a reduced set of channel candidates, which may lead to
increased fees. To let users choose whether the factor is active or not,
we add a config setting where a capacity fraction of 1.0 disables the
factor. We limit the capacity fraction to values between 0.75 and 1.0.
Lower values may discard too many channels.
Implements a new probability estimator based on a probability theory
framework.
The computed probability consists of:
* the direct channel probability, which is estimated based on a
depleted liquidity distribution model, formulas and broader concept derived
after Pickhardt et al. https://arxiv.org/abs/2103.08576
* an extension of the probability model to incorporate knowledge decay
after time for previous successes and failures
* a mixed node probability taking into account successes/failures on other
channels of the node (similar to the apriori approach)
* we rename the current probability estimator to be the "apriori"
probability estimator to distinguish from a different implementation
later
* the AprioriEstimator is exported to later be able to type switch
* getLocalPairProbability -> LocalPairProbabiltiy (later part of an
exported interface)
* getPairProbability -> getPairProbabiltiy (later part of an exported
interface)
