Update the evolutionary core to the fitness order, the selection inventory and a termination module - #110
Merged
Merged
Conversation
…ntory and a termination module Problem: The evolutionary package carries a driver, a fitness adapter and one selection method per role, and all of them read the fitness as a number. A search that compares several objectives has no place there: selection ranks by a scalar, the driver's best-so-far bookkeeping asks for the fittest member rather than for a member that beats the incumbent, and a termination condition is a closure the caller writes by hand. The components also state no contract about what a fitness value may be, so a measurement that failed and a measurement that is infinitely good travel through the same path. Changes: - The fitness of an individual is a value in a partially ordered set, and a comparator decides the order. ScalarFitnessComparator is the total order over reals, ParetoFitnessComparator the componentwise order over vectors. A scalarization maps into the strictly positive reals where a component draws proportionally, and only there. - The selection inventory covers the three roles: FitnessProportionalSelection, TournamentSelection and RankBasedSelection for parents, FitnessBasedReplacement and GenerousConservativeReplacement for survivors. dominance_fronts is the shared construction, the canonical linearization of a partial order into ranks. - RankBasedSelection keeps the ranking of the population it was last asked about. A driver asks for one pair per variation pass, every pass of a generation over the same population, and building the fronts is quadratic, so answering each pass from scratch multiplied that cost by the number of passes. The stored ranking answers again only if the population holds the same individuals in the same order and the fitness mapping and the comparator are the very objects of the previous call. On the shipped example this is 2.3 seconds instead of 37.6 for three generations of 250 individuals, measured on one machine in one session. - A new module termination with Generations, TargetFitness, NoImprovement and Any/All combinators, so a condition is a component rather than a closure. - The best-so-far is read from the parents and the offspring together rather than from the survivors. A survivor selection keeps mu individuals and is a parameter of the search, so under a partial order it may drop an offspring that beats the incumbent, and a dropped individual never gets a second chance to be reported. Among the members that do beat the incumbent a maximal one is taken, which keeps the incumbent monotone where a scan for one maximal element would not. - ExpScalarization reports a finite fitness that overflows the exponential instead of returning infinity. A proportional draw reads infinity as an infinitely good individual and gives it the whole mass. The underflowing end is reported by the drawing component, which already refuses a weight of zero. - The driver checks what its components return: the initializer for the population size, the survivor selection for the size and for closure over parents and offspring. - The shipped example runs 100 generations from seed 4 and reaches a training error of 1.3e-07, where the previous configuration kept the error of its initial population. Truncation over parents and offspring together gives every copy of an individual a place of its own, and a pass that neither recombines nor mutates hands a parent on unchanged, so the multiplicity of the fittest individuals grows and a run can reach a population that holds few distinct terms. Which seed does that is a property of the run rather than of the components. Behavior: - A seeded run does not reproduce its previous results. The selection methods are different components, and the driver's bookkeeping reads a different set. - evolutionary_best can return an offspring the survivor selection did not keep, and last_improvement can advance in a generation whose population did not improve, so NoImprovement stops such a run at a different point. - ExpScalarization raises where it returned infinity before.
![tudo-seal-workflows[bot]](https://avatars.githubusercontent.com/in/2312552?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Benchmark CoSy
Details
| Benchmark suite | Current: ef90aef | Previous: be71b5d | Ratio |
|---|---|---|---|
benchmarks/test_benchmark_maximal_elements.py::test_benchmark_maximal_elements |
9.535558257894504 iter/sec (stddev: 0.0004587768100170705) |
9.608281351562356 iter/sec (stddev: 0.010017085420616013) |
1.01 |
benchmarks/test_benchmark_maze.py::test_benchmark_maze |
3.5747029431505406 iter/sec (stddev: 0.02532070555591522) |
3.9235166334572096 iter/sec (stddev: 0.018797551923491352) |
1.10 |
benchmarks/test_benchmark_maze_contains.py::test_benchmark_maze_contains |
3.3518636219954354 iter/sec (stddev: 0.024308568544443042) |
3.5391214334611916 iter/sec (stddev: 0.025877608823890046) |
1.06 |
benchmarks/test_benchmark_maze_loopfree.py::test_benchmark_maze_loopfree |
3.700453691164463 iter/sec (stddev: 0.01563164705813299) |
3.8477221717914887 iter/sec (stddev: 0.02072136799063595) |
1.04 |
This comment was automatically generated by workflow using github-action-benchmark.
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## develop #110 +/- ##
===========================================
+ Coverage 88.76% 91.37% +2.60%
===========================================
Files 69 73 +4
Lines 7435 8165 +730
Branches 829 844 +15
===========================================
+ Hits 6600 7461 +861
+ Misses 739 627 -112
+ Partials 96 77 -19
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem:
The evolutionary package carries a driver, a fitness adapter, and one selection method per role, and
all of them read the fitness as a number. A search that compares several objectives has no place
there:
ParetoFitnessComparatorreports0for both a tie and an incomparable pair, itsscalarizesums the objectives, and every component that sorts reads that sum, so a Pareto run isa weighted-sum run under another name. The best-so-far bookkeeping asks for the fittest member
rather than for a member that beats the incumbent. A termination condition is a closure the caller
writes by hand, and nothing states what a fitness value may be, so a measurement that failed and one
that is infinitely good travel the same path.
Changes:
order.
compareanswersGREATER,LESS,EQUALorINCOMPARABLE, so a tie and anincomparable pair are no longer the same answer.
ScalarFitnessComparatoris the total orderover reals and reports
INCOMPARABLEfornan,ParetoFitnessComparatoris the componentwiseorder over vectors.
is passed explicitly to the components that draw proportionally, and only to those, so choosing
one is a visible decision rather than a summed default.
FitnessProportionalSelection,TournamentSelectionandRankBasedSelectionfor parents,FitnessBasedReplacementandGenerousConservativeReplacementfor survivors.dominance_frontsis the shared construction,the canonical linearization of a partial order into ranks.
RankBasedSelectionkeeps the ranking of the population it was last asked about. A driver asksfor one pair per variation pass, every pass of a generation over the same population, and
building the fronts is quadratic, so answering each pass from scratch multiplies that cost by the
number of passes. The stored ranking answers again only if the population holds the same
individuals in the same order and the fitness mapping and the comparator are the very objects of
the previous call. On the shipped example, this is 2.3 seconds instead of 37.6 for three
generations of 250 individuals, measured on one machine in one session.
terminationwithGenerations,TargetFitness,NoImprovementandAny/Allcombinators, so a condition is a component rather than a closure.
survivors. A survivor selection keeps
muindividuals and is a parameter of the search, so undera partial order it may drop an offspring that beats the incumbent, and a dropped individual never
gets a second chance to be reported. Among the members that do beat the incumbent, a maximal one
is taken, which keeps the incumbent monotone where a scan for one maximal element would not.
ExpScalarizationreports a finite fitness that overflows the exponential instead of returninginfinity. A proportional draw reads infinity as an infinitely good individual and gives it the
whole mass. The underflowing end is reported by the drawing component, which already refuses a
weight of zero.
survivor selection for the size and for closure over parents and offspring.
where the previous configuration kept the error of its initial population. Truncation over
parents and offspring together gives every copy of an individual a place of its own, and a pass
that neither recombines nor mutates hands a parent on unchanged, so the multiplicity of the
fittest individuals grows and a run can reach a population that holds few distinct terms. Which
seed is used is a property of the run rather than of the components.
SimpleGeneticProgrammingis replaced byEvolutionarySearch, andEvolutionary,SelectionandAgeBasedReplacementare gone.comparereturns aComparisonrather than anint, and a comparator no longer carriesscalarizeorsort_key.different components and the bookkeeping reads a different set.
evolutionary_bestcan return anoffspring the survivor selection did not keep, and
last_improvementcan advance in a generationwhose population did not improve, so
NoImprovementstops such a run at a different point.ExpScalarizationraises where it returned infinity before.