Implement MOEA/D - DE #269

jonpsy · 2021-03-09T10:28:26Z

This PR is a continuation of #210. Till now I've made these changes

Address issue #176

Naming changes
Fixing algorithms
Wrapping codes into functions
Argument passing fixed

**Important: For more details regarding its testing and comparison against other algorithms. Refer this discussion.

The following remains to be done

TODO

Adding more sanity checks
~~Let user explicitly choose whether they want to preserve diversity or not~~
~~Use a Policy-based approach to delegate Weight initialization and Decomposition.~~
Discard arma::mat for InputType and OutputType.
Add test cases for various MatType: Stress testing and accuracy measurement. Schaffer and Fonseca pass currently for
double.
Use the CheckArbitraryFunctionType API once that PR is merged.
Assert RequireDenseFloatingType InputType.
~~After ZDT is merged, test against it: report results. (or maybe do it in separate PR?)~~

FIXME

Handle zero weight condition in Tchebycheff decomposition
Updating Solutions

DOCUMENTATION

Write about MOEA/D-DE in optimizer.md
Update History.MD
Finalize docs.

Memory

Find and remove redundant & expensive matrix copy.
Find and remove leaks

Testing

Add callback test

REVIEWS
Requesting @zoq @say4n @rcurtin for reviews.

jonpsy · 2021-03-09T10:37:55Z

There are a few compile-time errors, I'll be fixing that over-the-course. I'd like reviews on algorithmic correctness and semantics.

jonpsy · 2021-03-10T21:20:28Z

Just have to fix the tests, a little bit of tinkering aaaand we should be done! I'd like to add other decomposition methods (Bi, Weighted) as part of separated PR so as to avoid bloating :)

jonpsy · 2021-03-11T19:01:47Z

Good news:
=> Everything compiles
Bad news:
=> All the test fail
=> preserveDiversity when set to false leads to SIG memory violation.

@zoq, the actual paper preserves diversity by default, but pagmo2 allows user to turn off that feature. By "preserving diversity", I mean to say the optimizer produces a diverse set of population.

It does by the following way (as per the paper):
a) Algorithm chooses to sample from either the neighbor or whole population, but favors the "whole population" thus increasing diversity.
b) The child solutions (however good they might be) cannot replace a threshold number of parents. For ex: if maxReplace = 10, no more than 10 children would replace original solution.

The above-mentioned feature can be switched off by disabling preserveDiversity, but I was wondering, is it overkill? Because paper didn't allow any choice and simply favored preserving diversity.

jonpsy · 2021-03-12T17:43:31Z

All tests pass! @zoq could you review this please? I was wondering we could try and compare NSGA2 & MOEA/D-DE in terms of speed, correctness.

jonpsy · 2021-03-13T18:29:06Z

Pushing some more. I'll be explaining my choice of default args over the next few comments.

zoq

Thanks, I'll take a closer look into the code over the next days, do you mind to go over the code and adjust the indentation according to the style guidelines.

doc/optimizers.md

include/ensmallen_bits/moead/moead.hpp

jonpsy · 2021-03-14T06:57:34Z

Default args:

populationSize: 150. For 2 objectives he's chosen N=300, so I took N=150, assuming single objective.

neighborSize : 20
neighborProb: 0.9
maxReplace: 2

Some notes: The paper recommends maxReplace << neighborSize

Finally

crossOverProb: 1.0
differentialWeight: 0.5
distributionIndex: 20
mutationRate: 1/numVariables

jonpsy · 2021-03-16T03:27:57Z

@zoq Gentle nudge :)

zoq

I have to take another look at the paper, but this looks really good.

include/ensmallen_bits/moead/moead.hpp

include/ensmallen_bits/moead/moead_impl.hpp

tests/moead_test.cpp

jonpsy · 2021-03-18T03:23:11Z

I was wondering why the indentation messes up when I push it here. Looks like tabs vs space issue, thanks for pointing them out @zoq.

jonpsy · 2021-04-10T16:54:58Z

Stalling this for a while, till NSGA2, Recursive MOO and Indicators gets merged. They will serve as a baseline for this pull request. Have updated TODO list for others to track the progress of this PR.

jonpsy · 2021-04-18T17:29:31Z

Keep open!

- populationSize > neighborSize + 1 - fix initNeighborhood - neighborHoodSize => neighborSize

- declare DifferentialCrossover

fix callback test

jonpsy · 2021-06-12T05:15:17Z

@zoq I think I've addressed all of your comments. If you find some indent mistakes, it would be helpful if you could do a "commit suggestion", it gets tiring to change minor things every commits. Thanks a lot.

jonpsy · 2021-06-12T07:18:04Z

Uhm... I've done a sizeable amount of work on the policy-based method already. How about we merge this as-is now, and create a follow-up PR in which I handle the templatizing and some other minor fixes.

Here's a patch for policy if you want a look.

zoq · 2021-06-12T17:15:51Z

@zoq I think I've addressed all of your comments. If you find some indent mistakes, it would be helpful if you could do a "commit suggestion", it gets tiring to change minor things every commits. Thanks a lot.

Maybe you can adjust your code editor to use the correct indentation style, I use vim and setting up the indentation is easy, That could save us both some time. https://github.com/mlpack/mlpack/wiki/DesignGuidelines has all the important style guidelines.

zoq · 2021-06-12T17:19:41Z

Uhm... I've done a sizeable amount of work on the policy-based method already. How about we merge this as-is now, and create a follow-up PR in which I handle the templatizing and some other minor fixes.

Here's a patch for policy if you want a look.

You can just open another PR with the changes from this one and the policy design changes, and rebase the PR afterwards. I don't want to rush anything. Like one thing I haven't done is run the tests in a loop and check if they are stable, maybe you did that already in which case we can check that one.

zoq

Next step to test this in a loop.

include/ensmallen_bits/moead/moead_impl.hpp

tests/moead_test.cpp

include/ensmallen_bits/moead/moead_impl.hpp

include/ensmallen_bits/moead/moead.hpp

Co-authored-by: Marcus Edel <marcus.edel@fu-berlin.de>

jonpsy · 2021-06-13T03:51:20Z

Thank you for taking the time to fix the style. I'll be sure to change my IDE settings.

Reg testing, one thing you'll find interesting to know is, that higher population size leads to more accurate results.

I have yet to test this claim, I'll play around some parameters run it over a 1000 times and report the results here.

zoq · 2021-06-13T17:31:51Z

Reg testing, one thing you'll find interesting to know is, that higher population size leads to more accurate results.

That is true, we basically increase the number of possible solutions. For the tests we don't need a perfect solution, just something that is close, to check that the optimizer works with different starting solutions.

jonpsy · 2021-06-14T10:17:50Z

Ran these command ctest -C Release -V --output-on-failure --repeat-until-fail 1000 --rng-seed time to test over 1000 times. It passed each and every time. link. Feel free (and I encourage you to) try this yourself. Just comment out other tests from CMakeLists.txt.

Let's do some final cleanups. Doc fixes, if any, and then this should be good to go imo. Looking forward for your review.

zoq

Great work, no further comments from my side.

mlpack-bot

Second approval provided automatically after 24 hours. 👍

jonpsy · 2021-06-15T16:45:41Z

All right, if nobody has any objections. Let's get this merged!

zoq · 2021-06-15T19:07:57Z

Nice, with that one merged, let's open another PR for the policy design changes.

mlpack-bot bot added s: needs review s: unanswered s: unlabeled labels Mar 9, 2021

jonpsy changed the title ~~Implement MOEA/D - DE~~ [WIP] Implement MOEA/D - DE Mar 9, 2021

zoq added c: optimizers and removed s: unanswered s: unlabeled labels Mar 9, 2021

jonpsy changed the title ~~[WIP] Implement MOEA/D - DE~~ Implement MOEA/D - DE Mar 12, 2021

jonpsy force-pushed the moead branch from 0b54a17 to da75ea7 Compare March 12, 2021 19:56

zoq reviewed Mar 13, 2021

View reviewed changes

doc/optimizers.md Outdated Show resolved Hide resolved

doc/optimizers.md Outdated Show resolved Hide resolved

doc/optimizers.md Outdated Show resolved Hide resolved

include/ensmallen_bits/moead/moead.hpp Outdated Show resolved Hide resolved

jonpsy force-pushed the moead branch from 1e392fa to aeddb53 Compare March 14, 2021 06:47

zoq reviewed Mar 18, 2021

View reviewed changes

say4n added the s: keep open label Apr 18, 2021

jonpsy added 7 commits June 8, 2021 10:20

- Import previous work

0419469

- populationSize > neighborSize + 1 - fix initNeighborhood - neighborHoodSize => neighborSize

- wrap mating in a function

19425d4

- declare DifferentialCrossover

crossover inside mainloop

91df706

differential crossover

3d91ccc

cast to vector only at EvaluateObjectives

a52f1ad

convention

bb414e6

scalingFactor getter setter

3cd94ae

fix default arg in ctor

53cae18

fix callback test

weights are fixed bruh

e893dfd

zoq reviewed Jun 12, 2021

View reviewed changes

jonpsy and others added 7 commits June 13, 2021 09:15

Update include/ensmallen_bits/moead/moead_impl.hpp

b1fa440

Co-authored-by: Marcus Edel <marcus.edel@fu-berlin.de>

Update include/ensmallen_bits/moead/moead_impl.hpp

a2a7ef1

Co-authored-by: Marcus Edel <marcus.edel@fu-berlin.de>

Update include/ensmallen_bits/moead/moead_impl.hpp

7d794cb

Co-authored-by: Marcus Edel <marcus.edel@fu-berlin.de>

Update include/ensmallen_bits/moead/moead_impl.hpp

40e44d4

Co-authored-by: Marcus Edel <marcus.edel@fu-berlin.de>

Update tests/moead_test.cpp

5cd6bc8

Co-authored-by: Marcus Edel <marcus.edel@fu-berlin.de>

Update include/ensmallen_bits/moead/moead_impl.hpp

da7683b

Co-authored-by: Marcus Edel <marcus.edel@fu-berlin.de>

Update include/ensmallen_bits/moead/moead.hpp

973d4a7

Co-authored-by: Marcus Edel <marcus.edel@fu-berlin.de>

zoq approved these changes Jun 14, 2021

View reviewed changes

mlpack-bot bot approved these changes Jun 15, 2021

View reviewed changes

mlpack-bot bot removed the s: needs review label Jun 15, 2021

zoq merged commit 3497f27 into mlpack:master Jun 15, 2021

jonpsy deleted the moead branch June 16, 2021 02:03

This was referenced Jun 16, 2021

Introduce Policy Methods for MOEA/D-DE #293

Merged

MOEA/D-DE vs NSGA-II : Battle of the Optimizers #294

Closed

This was referenced Jul 5, 2021

Release version 2.17.0: "25 Din Mein Paisa Double" #308

Closed

Release version 2.17.0: "Pachis Din Me Pesa Double" #310

Merged

mlpack deleted a comment from shikharjan Nov 29, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement MOEA/D - DE #269

Implement MOEA/D - DE #269

jonpsy commented Mar 9, 2021 •

edited

Loading

jonpsy commented Mar 9, 2021

jonpsy commented Mar 10, 2021

jonpsy commented Mar 11, 2021

jonpsy commented Mar 12, 2021

jonpsy commented Mar 13, 2021

zoq left a comment

jonpsy commented Mar 14, 2021

jonpsy commented Mar 16, 2021

zoq left a comment

jonpsy commented Mar 18, 2021

jonpsy commented Apr 10, 2021 •

edited

Loading

jonpsy commented Apr 18, 2021 •

edited

Loading

jonpsy commented Jun 12, 2021

jonpsy commented Jun 12, 2021 •

edited

Loading

zoq commented Jun 12, 2021

zoq commented Jun 12, 2021

zoq left a comment

jonpsy commented Jun 13, 2021

zoq commented Jun 13, 2021

jonpsy commented Jun 14, 2021

zoq left a comment

mlpack-bot bot left a comment

jonpsy commented Jun 15, 2021

zoq commented Jun 15, 2021

Implement MOEA/D - DE #269

Implement MOEA/D - DE #269

Conversation

jonpsy commented Mar 9, 2021 • edited Loading

jonpsy commented Mar 9, 2021

jonpsy commented Mar 10, 2021

jonpsy commented Mar 11, 2021

jonpsy commented Mar 12, 2021

jonpsy commented Mar 13, 2021

zoq left a comment

Choose a reason for hiding this comment

jonpsy commented Mar 14, 2021

jonpsy commented Mar 16, 2021

zoq left a comment

Choose a reason for hiding this comment

jonpsy commented Mar 18, 2021

jonpsy commented Apr 10, 2021 • edited Loading

jonpsy commented Apr 18, 2021 • edited Loading

jonpsy commented Jun 12, 2021

jonpsy commented Jun 12, 2021 • edited Loading

zoq commented Jun 12, 2021

zoq commented Jun 12, 2021

zoq left a comment

Choose a reason for hiding this comment

jonpsy commented Jun 13, 2021

zoq commented Jun 13, 2021

jonpsy commented Jun 14, 2021

zoq left a comment

Choose a reason for hiding this comment

mlpack-bot bot left a comment

Choose a reason for hiding this comment

jonpsy commented Jun 15, 2021

zoq commented Jun 15, 2021

jonpsy commented Mar 9, 2021 •

edited

Loading

jonpsy commented Apr 10, 2021 •

edited

Loading

jonpsy commented Apr 18, 2021 •

edited

Loading

jonpsy commented Jun 12, 2021 •

edited

Loading