Files
pezkuwi-subxt/substrate/primitives/npos-elections/src/phragmen.rs
T
Peter Goodspeed-Niklaus 781f908760 Implement PJR checker (#8160)
* Apply.

* get rid of glob import

* use meaningful generic type name

* pjr_check operates on `Supports` struct used elsewhere

* improve algorithmic complexity of `prepare_pjr_input`

* fix rustdoc warnings

* improve module docs

* typo

* simplify debug assertion

* add test finding the phase-change threshold value for a constructed scenario

* add more threshold scenarios to disambiguate plausible interpretations

* add link to npos paper reference

* docs: staked_assignment -> supports

Co-authored-by: Kian Paimani <5588131+kianenigma@users.noreply.github.com>

* add utility method for generating npos inputs

* add a fuzzer which asserts that all unbalanced seq_phragmen are PJR

Note that this currently fails. I hope that this can be rectified
by calculating the threshold instead of choosing some arbitrary number.

* assert in all cases, not just debug

* leverage a native solution to choose candidates

* use existing helper methods

* add pjr-check and incorporate into the fuzzer

We should probably have one of the W3F people look at this to ensure
we're not misconstruing any definitions, but this seems like a
fairly straightforward implementation.

* fix compilation errors

* Enable manually setting iteration parameters in single run.

This gives us the ability to reproducably extract cases where
honggfuzz has discovered a panic. For example:

$ cargo run --release --bin phragmen_pjr -- --candidates 569 --voters 100
Tue 23 Feb 2021 11:23:39 AM CET
   Compiling bitflags v1.2.1
   Compiling unicode-width v0.1.8
   Compiling unicode-segmentation v1.7.1
   Compiling ansi_term v0.11.0
   Compiling strsim v0.8.0
   Compiling vec_map v0.8.2
   Compiling proc-macro-error-attr v1.0.4
   Compiling proc-macro-error v1.0.4
   Compiling textwrap v0.11.0
   Compiling atty v0.2.14
   Compiling heck v0.3.2
   Compiling clap v2.33.3
   Compiling structopt-derive v0.4.14
   Compiling structopt v0.3.21
   Compiling sp-npos-elections-fuzzer v2.0.0-alpha.5 (/home/coriolinus/Documents/Projects/paritytech/substrate/primitives/npos-elections/fuzzer)
    Finished release [optimized] target(s) in 6.15s
     Running `/home/coriolinus/Documents/Projects/paritytech/substrate/target/release/phragmen_pjr -c 569 -v 100`
thread 'main' panicked at 'unbalanced sequential phragmen must satisfy PJR', primitives/npos-elections/fuzzer/src/phragmen_pjr.rs:133:5
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

This is still not adequate proof that seq_phragmen is broken; it could
very well be that our PJR checker is doing the wrong thing, or we've
somehow missed a parameter of interest. Still, it's concerning.

* update comment verbiage for accuracy

* it is valid in PJR for an elected candidate to have 0 support

* Fix phragmen_pjr fuzzer

It turns out that the fundamental problem causing previous implementations
of the fuzzer to fail wasn't in `seq_phragmen` _or_ in `pjr_check`: it was
in the rounding errors introduced in the various conversions between the
internal data representation and the external one.

Fixing the fuzzer is then simply an issue of using the internal representation
and staying in that representation. However, that leaves the issue that
`seq_phragmen` occasionally produces an output which is technically not
PJR due to rounding errors. In the future we will need to add some kind of
"close-enough" threshold. However, that is explicitly out of scope of
this PR.

* restart ci; it appears to be stalled

* use necessary import for no-std

* use a more realistic distribution of voters and candidates

This isn't ideal; more realistic numbers would be about twice these.
However, either case generation or voting has nonlinear execution
time, and doubling these values brings iteration time from ~20s to
~180s. Fuzzing 6x as fast should make up for fuzzing cases half the size.

* identify specifically which PJR check may fail

* move candidate collection comment into correct place

* standard_threshold: use a calculation method which cannot overflow

* Apply suggestions from code review (update comments)

Co-authored-by: Kian Paimani <5588131+kianenigma@users.noreply.github.com>

* clarify the effectiveness bounds for t-pjr check

* how to spell "committee"

* reorganize: high -> low abstraction

* ensure standard threshold calc cannot panic

Co-authored-by: Kian Paimani <5588131+kianenigma@users.noreply.github.com>

* Apply suggestions from code review

Co-authored-by: Shawn Tabrizi <shawntabrizi@gmail.com>

Co-authored-by: kianenigma <kian.peymani@gmail.com>
Co-authored-by: Kian Paimani <5588131+kianenigma@users.noreply.github.com>
Co-authored-by: Shawn Tabrizi <shawntabrizi@gmail.com>
2021-03-11 09:06:53 +00:00

215 lines
8.2 KiB
Rust

// This file is part of Substrate.
// Copyright (C) 2020-2021 Parity Technologies (UK) Ltd.
// SPDX-License-Identifier: Apache-2.0
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//! Implementation of the sequential-phragmen election method.
//!
//! This method is ensured to achieve PJR, yet, it does not achieve a constant factor approximation
//! to the Maximin problem.
use crate::{
balancing, setup_inputs, CandidatePtr, ElectionResult, ExtendedBalance, IdentifierT,
PerThing128, VoteWeight, Voter,
};
use sp_arithmetic::{
helpers_128bit::multiply_by_rational,
traits::{Bounded, Zero},
Rational128,
};
use sp_std::prelude::*;
/// The denominator used for loads. Since votes are collected as u64, the smallest ratio that we
/// might collect is `1/approval_stake` where approval stake is the sum of votes. Hence, some number
/// bigger than u64::max_value() is needed. For maximum accuracy we simply use u128;
const DEN: ExtendedBalance = ExtendedBalance::max_value();
/// Execute sequential phragmen with potentially some rounds of `balancing`. The return type is list
/// of winners and a weight distribution vector of all voters who contribute to the winners.
///
/// - This function is a best effort to elect `rounds` members. Nonetheless, if less candidates are
/// available, it will only return what is available. It is the responsibility of the call site to
/// ensure they have provided enough members.
/// - If `balance` parameter is `Some(i, t)`, `i` iterations of balancing is with tolerance `t` is
/// performed.
/// - Returning winners are sorted based on desirability. Voters are unsorted. Nonetheless,
/// seq-phragmen is in general an un-ranked election and the desirability should not be
/// interpreted with any significance.
/// - The returning winners are zipped with their final backing stake. Yet, to get the exact final
/// weight distribution from the winner's point of view, one needs to build a support map. See
/// [`crate::SupportMap`] for more info. Note that this backing stake is computed in
/// ExtendedBalance and may be slightly different that what will be computed from the support map,
/// due to accuracy loss.
/// - The accuracy of the returning edge weight ratios can be configured via the `P` generic
/// argument.
/// - The returning weight distribution is _normalized_, meaning that it is guaranteed that the sum
/// of the ratios in each voter's distribution sums up to exactly `P::one()`.
///
/// This can only fail of the normalization fails. This can happen if for any of the resulting
/// assignments, `assignment.distribution.map(|p| p.deconstruct()).sum()` fails to fit inside
/// `UpperOf<P>`. A user of this crate may statically assert that this can never happen and safely
/// `expect` this to return `Ok`.
///
/// This can only fail if the normalization fails.
///
/// Note that rounding errors can potentially cause the output of this function to fail a t-PJR
/// check where t is the standard threshold. The underlying algorithm is sound, but the conversions
/// between numeric types can be lossy.
pub fn seq_phragmen<AccountId: IdentifierT, P: PerThing128>(
rounds: usize,
initial_candidates: Vec<AccountId>,
initial_voters: Vec<(AccountId, VoteWeight, Vec<AccountId>)>,
balance: Option<(usize, ExtendedBalance)>,
) -> Result<ElectionResult<AccountId, P>, crate::Error> {
let (candidates, voters) = setup_inputs(initial_candidates, initial_voters);
let (candidates, mut voters) = seq_phragmen_core::<AccountId>(
rounds,
candidates,
voters,
)?;
if let Some((iterations, tolerance)) = balance {
// NOTE: might create zero-edges, but we will strip them again when we convert voter into
// assignment.
let _iters = balancing::balance::<AccountId>(&mut voters, iterations, tolerance);
}
let mut winners = candidates
.into_iter()
.filter(|c_ptr| c_ptr.borrow().elected)
// defensive only: seq-phragmen-core returns only up to rounds.
.take(rounds)
.collect::<Vec<_>>();
// sort winners based on desirability.
winners.sort_by_key(|c_ptr| c_ptr.borrow().round);
let mut assignments =
voters.into_iter().filter_map(|v| v.into_assignment()).collect::<Vec<_>>();
let _ = assignments
.iter_mut()
.map(|a| a.try_normalize().map_err(|e| crate::Error::ArithmeticError(e)))
.collect::<Result<(), _>>()?;
let winners = winners
.into_iter()
.map(|w_ptr| (w_ptr.borrow().who.clone(), w_ptr.borrow().backed_stake))
.collect();
Ok(ElectionResult { winners, assignments })
}
/// Core implementation of seq-phragmen.
///
/// This is the internal implementation that works with the types defined in this crate. see
/// `seq_phragmen` for more information. This function is left public in case a crate needs to use
/// the implementation in a custom way.
///
/// This can only fail if the normalization fails.
// To create the inputs needed for this function, see [`crate::setup_inputs`].
pub fn seq_phragmen_core<AccountId: IdentifierT>(
rounds: usize,
candidates: Vec<CandidatePtr<AccountId>>,
mut voters: Vec<Voter<AccountId>>,
) -> Result<(Vec<CandidatePtr<AccountId>>, Vec<Voter<AccountId>>), crate::Error> {
// we have already checked that we have more candidates than minimum_candidate_count.
let to_elect = rounds.min(candidates.len());
// main election loop
for round in 0..to_elect {
// loop 1: initialize score
for c_ptr in &candidates {
let mut candidate = c_ptr.borrow_mut();
if !candidate.elected {
// 1 / approval_stake == (DEN / approval_stake) / DEN. If approval_stake is zero,
// then the ratio should be as large as possible, essentially `infinity`.
if candidate.approval_stake.is_zero() {
candidate.score = Bounded::max_value();
} else {
candidate.score = Rational128::from(DEN / candidate.approval_stake, DEN);
}
}
}
// loop 2: increment score
for voter in &voters {
for edge in &voter.edges {
let mut candidate = edge.candidate.borrow_mut();
if !candidate.elected && !candidate.approval_stake.is_zero() {
let temp_n = multiply_by_rational(
voter.load.n(),
voter.budget,
candidate.approval_stake,
).unwrap_or(Bounded::max_value());
let temp_d = voter.load.d();
let temp = Rational128::from(temp_n, temp_d);
candidate.score = candidate.score.lazy_saturating_add(temp);
}
}
}
// loop 3: find the best
if let Some(winner_ptr) = candidates
.iter()
.filter(|c| !c.borrow().elected)
.min_by_key(|c| c.borrow().score)
{
let mut winner = winner_ptr.borrow_mut();
// loop 3: update voter and edge load
winner.elected = true;
winner.round = round;
for voter in &mut voters {
for edge in &mut voter.edges {
if edge.who == winner.who {
edge.load = winner.score.lazy_saturating_sub(voter.load);
voter.load = winner.score;
}
}
}
} else {
break
}
}
// update backing stake of candidates and voters
for voter in &mut voters {
for edge in &mut voter.edges {
if edge.candidate.borrow().elected {
// update internal state.
edge.weight = multiply_by_rational(
voter.budget,
edge.load.n(),
voter.load.n(),
)
// If result cannot fit in u128. Not much we can do about it.
.unwrap_or(Bounded::max_value());
} else {
edge.weight = 0
}
let mut candidate = edge.candidate.borrow_mut();
candidate.backed_stake = candidate.backed_stake.saturating_add(edge.weight);
}
// remove all zero edges. These can become phantom edges during normalization.
voter.edges.retain(|e| e.weight > 0);
// edge of all candidates that eventually have a non-zero weight must be elected.
debug_assert!(voter.edges.iter().all(|e| e.candidate.borrow().elected));
// inc budget to sum the budget.
voter.try_normalize_elected().map_err(|e| crate::Error::ArithmeticError(e))?;
}
Ok((candidates, voters))
}