Way to `teach' NSGA-II with previous data?

[jbae11]

• inspyred version: 1.0
• Python version: 3.7.3
• Operating System: MAC OS

Description

I am using Inspyred to run NSGA-II to do multiobjective optimization on a very computationally heavy problem. While running, my cluster crashed and I lost that run. I have 40 populations (samples and results), but it seems like I have to start all over again. Is there a way to 'feed' previously generated data into NSGA-II so that it starts off with that knowledge?

Thanks for the great tool!

[aarongarrett]

Yes. You can pass a list of candidates as seeds into the evolve method. For instance, suppose you had two candidates that you wanted to start from: [0, 8, 6, 7] and [5, 3, 0, 9]. Then you could make a call like this to evolve: myec.evolve(generator=mygen, evaluator=myeval, pop_size=100, seeds=[[0, 8, 6, 7], [5, 3, 0, 9]]) Obviously, if you have lots of stuff, you'd prefer to read them from a file and build that list programmatically. Also, make sure your pop_size is at least as large as your number of seed candidates.

…

-- Aaron Garrett

On Mon, Aug 12, 2019 at 1:50 PM Jin Whan Bae ***@***.***> wrote: - inspyred version: 1.0 - Python version: 3.7.3 - Operating System: MAC OS Description I am using Inspyred to run NSGA-II to do multiobjective optimization on a very computationally heavy problem. While running, my cluster crashed and I lost that run. I have 40 populations (samples and results), but it seems like I have to start all over again. Is there a way to 'feed' previously generated data into NSGA-II so that it starts off with that knowledge? Thanks for the great tool! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#18?email_source=notifications&email_token=AALRPXMO6QGWMJG3OZJPAADQEGPGJA5CNFSM4ILDRKJ2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HEYXJLA>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AALRPXL7XNMS7PA2FROSUU3QEGPGJANCNFSM4ILDRKJQ> .

[jbae11]

thanks for the very quick reply. Is it possible to maybe give them the 'solutions' as well?
For example, if I am optimizing a multiobjective function with 4 inputs and 2 outputs, can I input 10 sets of 4 inputs and 2 outputs so that the NSGA-II starts with those datasets 'in mind'?

[aarongarrett]

So you also want to give it the fitness values? That's not as simple to do. We'd have to write some custom code for that. Or maybe you could load a memoized wrapper with the fitness values for your seed candidates so that the evaluator wouldn't really have to run on those...

…

-- Aaron Garrett

On Mon, Aug 12, 2019 at 2:41 PM Jin Whan Bae ***@***.***> wrote: thanks for the very quick reply. Is it possible to maybe give them the 'solutions' as well? For example, if I am optimizing a multiobjective function with 4 inputs and 2 outputs, can I input 10 sets of 4 inputs and 2 outputs so that the NSGA-II starts with those datasets 'in mind'? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#18?email_source=notifications&email_token=AALRPXMR5IAMMYHGDQAFDHDQEGVEDA5CNFSM4ILDRKJ2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4DOOFY#issuecomment-520546071>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AALRPXPTO72MUIVTHAWFWALQEGVEDANCNFSM4ILDRKJQ> .

[jbae11]

I see, so the best way would be to give them my best candidates for the first generation?

[aarongarrett]

Well, we could figure out a way to get the fitness values loaded, but it would take some coding. I guess it depends on how long your evaluations take. If they only take 30 minutes total to re-evaluate, then that would be far faster than working out a programmatic solution to your problem.

…

-- Aaron Garrett

On Mon, Aug 12, 2019 at 2:45 PM Jin Whan Bae ***@***.***> wrote: I see, so the best way would be to give them my best candidates for the first generation? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#18?email_source=notifications&email_token=AALRPXO3W7OAH7SOTKHE5I3QEGVTBA5CNFSM4ILDRKJ2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4DOY7Q#issuecomment-520547454>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AALRPXNURRYGQVALDGQLEDLQEGVTBANCNFSM4ILDRKJQ> .

[jbae11]

My evaluations take a couple of hours per single evaluation. But I wouldn't want to trouble you with coming up with a new developments.. I bet my use case is quite rare..?

[aarongarrett]

Are you able to send me your code to look at? If so, send it to aaron.lee.garrett@gmail.com

…

-- Aaron Garrett

On Mon, Aug 12, 2019 at 2:53 PM Jin Whan Bae ***@***.***> wrote: My evaluations take a couple of hours per single evaluation. But I wouldn't want to trouble you with coming up with a new developments.. I bet my use case is quite rare..? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#18?email_source=notifications&email_token=AALRPXLY743GXQHOCB76XJ3QEGWUBA5CNFSM4ILDRKJ2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4DPSOQ#issuecomment-520550714>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AALRPXMIJMHZEC537ID6UGTQEGWUBANCNFSM4ILDRKJQ> .

[jbae11]

I'm afraid not, it's not quite for public release. I'm sorry

[jbae11]

Also, can you tell me what's the best way to cite your software? thanks

[aarongarrett]

Sure. This is what others have done: https://groups.google.com/d/msg/inspyred/K43UQsCLitM/7ltTXyDbAQAJ

…

-- Aaron Garrett

On Mon, Aug 12, 2019 at 3:05 PM Jin Whan Bae ***@***.***> wrote: Also, can you tell me what's the best way to cite your software? thanks — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#18?email_source=notifications&email_token=AALRPXM63AUTJAJRGT2SIW3QEGX7TA5CNFSM4ILDRKJ2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4DQSQI#issuecomment-520554817>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AALRPXKRLYI6NLFYCWLSLJLQEGX7TANCNFSM4ILDRKJQ> .