Dear all,
I have implemented an OpenTURNSPythonFunction <http://www.simxon.com/articles/20190124-cae-statistics-and-robust-design.ht ml> which can do Python-style multiprocessing when invoked by an exec_sample(<list>)-call. This means that OpenTURNS will under these circumstances not even know that multiprocessing takes place. I guess that Julien has answered my question to the extent that I need. If I implement CONDOR soon, I will not interfere when anyone elses work. Would anyone care to join me for discussions or whatever? Regards, Kim Ravn-Jensen Simxon (Danish VAT No.: 2685 6337) Phone: +45 2145 8811 <http://www.simxon.com/> http://www.simxon.com Fra: [email protected] <[email protected]> På vegne af Julien Schueller | Phimeca Sendt: 25. juni 2019 14:21 Til: 'OpenTURNS users forum' <[email protected]> Emne: Re: [ot-users] CONDOR vs. COBYLA? Hello Kim, Welcome to OpenTURNS. You're right COBYLA is sequential, and OpenTURNS evaluates in batches when using finite-difference for the gradients, so we cannot fully benefit from parallel evaluations. We're currently working on wiring more optimization libraries, but not specifically CONDOR, nor specifically oriented towards parallel computing. However we'll consider any contributions and feedback on this. https://www.sciencedirect.com/science/article/pii/S037704270400559X (I tried grabbing the code but the link in the article is dead) Also this may be a personal view, but cobyla is rather ancient, maybe there exist some are more modern algorithms (cmaes, nsga2) that can parallelize the optimization, we havent looked at that much, except for the multiple-objective aspect. j _____ De : [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]> > de la part de Simxon - Kim Ravn-Jensen <[email protected] <mailto:[email protected]> > Envoyé : dimanche 23 juin 2019 08:11:55 À : 'OpenTURNS users forum' Objet : [ot-users] CONDOR vs. COBYLA? Dear all, Even though I am a complete newbie concerning OpenTURNS, I recently took a chance and wrote <http://smtp.datacenter.ibo.fr:32224/?dmVyPTEuMDAxJiZkOTkwYWI3MzBhOTIxOTI0OT 01RDEyMDhFMF81MjE4Ml8zNTEwXzEmJjliNmVjZjUwYjgzODk4Nj0xMjMzJiZ1cmw9aHR0cCUzQS UyRiUyRnd3dyUyRXNpbXhvbiUyRWNvbSUyRmFydGljbGVzJTJGMjAxOTAxMjQtY2FlLXN0YXRpc3 RpY3MtYW5kLXJvYnVzdC1kZXNpZ24lMkVodG1s> this article. While researching, I noticed that the COBYLA algorithm uses parallel computing only when calculating gradients and Hessians for the destination point. If you google these two words together: COBYLA CONDOR, you are led to believe that the CONDOR algorithm is a possible replacement for COBYLA which is able to operate in parallel before reaching the final destination the optimum point. This leads to three questions for the user forum: 1. Have I got the facts right? 2. Assuming a yes to item 1, do you see a potential improvement of OpenTURNS in enabling support for CONDOR as an alternative to COBYLA? 3. Are any of you already working on it? Regards, Kim Ravn-Jensen Simxon (Danish VAT No.: 2685 6337) Phone: +45 2145 8811 <http://smtp.datacenter.ibo.fr:32224/?dmVyPTEuMDAxJiY5Zjk3YjY2ZTQ4ZDI1ODdkOD 01RDEyMDhFMF81MjE4Ml8zNTEwXzEmJmRhZWFmYTkwMzljODU4MD0xMjMzJiZ1cmw9aHR0cCUzQS UyRiUyRnd3dyUyRXNpbXhvbiUyRWNvbSUyRg==> http://www.simxon.com .
_______________________________________________ OpenTURNS users mailing list [email protected] http://openturns.org/mailman/listinfo/users
