MATLAB and Simulink resources for Arduino, LEGO, and Raspberry Pi
Learn moreOpportunities for recent engineering grads.
Apply TodayMATLAB Central > MATLAB Newsreader > Scaling problem with fmincon 

Subject: Scaling problem with fmincon From: Sargondjani Date: 26 Jun, 2012 11:34:06 Message: 1 of 24 
hi, i want to maximize the expected (infinite time) utility. There are X number of states, and the choice variables are consumption (C) in every state. 
Subject: Scaling problem with fmincon From: Steve Grikschat Date: 26 Jun, 2012 14:11:07 Message: 2 of 24 
"Sargondjani" wrote in message <jsc6ne$csf$1@newscl01ah.mathworks.com>... 
"Sargondjani" wrote in message <jsc6ne$csf$1@newscl01ah.mathworks.com>... 
Subject: Scaling problem with fmincon From: Sargondjani Date: 26 Jun, 2012 20:52:07 Message: 4 of 24 
"Steve Grikschat" wrote in message <jscftr$p6u$1@newscl01ah.mathworks.com>... 
Subject: Scaling problem with fmincon From: Sargondjani Date: 26 Jun, 2012 21:03:06 Message: 5 of 24 
The whole thing is differentiable, so that's not the problem... 
"Sargondjani" wrote in message <jsd82a$ijj$1@newscl01ah.mathworks.com>... 
Subject: Scaling problem with fmincon From: Sargondjani Date: 14 Jul, 2012 10:04:27 Message: 7 of 24 
> What I'm saying is that you haven't defined "accurate" clearly enough. If there are two points C1 and C2 which optimize E_V, then you should consider them both accurate. 
"Sargondjani" wrote in message <jtrg7b$bef$1@newscl01ah.mathworks.com>... 
Subject: Scaling problem with fmincon From: Sargondjani Date: 16 Jul, 2012 10:17:33 Message: 9 of 24 
"Matt J" wrote in message <jtse8k$mdq$1@newscl01ah.mathworks.com>... 
"Sargondjani" wrote in message <ju0pnt$a75$1@newscl01ah.mathworks.com>... 
Subject: Scaling problem with fmincon From: Sargondjani Date: 16 Jul, 2012 12:27:13 Message: 11 of 24 
> I doubt it, but I'll be interested to here if it works. 
"Sargondjani" wrote in message <ju11b1$7ni$1@newscl01ah.mathworks.com>... 
Subject: Scaling problem with fmincon From: Sargondjani Date: 17 Jul, 2012 16:19:25 Message: 13 of 24 
> > Hmmmm, I see. I still wonder how well the 'internal scaling' works, because in the other thread you also gave an example of bad scaling (one variable being 1e6 times larger) 
Subject: Scaling problem with fmincon From: Sargondjani Date: 18 Jul, 2012 07:28:13 Message: 14 of 24 
i have run some additional tests and if i just look at the conditional number, i found that scaling so the gradient gets of similar magnitude does not help with sqp, but sometimes with active set (and even then setting typicalX to 1/gradient actually helped more) 
Subject: Scaling problem with fmincon From: Bruno Luong Date: 18 Jul, 2012 08:16:19 Message: 15 of 24 
"Sargondjani" wrote in message <ju5oid$nk$1@newscl01ah.mathworks.com>... 
"Sargondjani" wrote in message <ju5oid$nk$1@newscl01ah.mathworks.com>... 
Subject: Scaling problem with fmincon From: Sargondjani Date: 21 Jul, 2012 19:34:26 Message: 17 of 24 
Thanks alot Bruno and Matt.... I think i understand alot better! 
"Sargondjani" wrote in message <juf082$gcm$1@newscl01ah.mathworks.com>... 
Subject: Scaling problem with fmincon From: Bruno Luong Date: 22 Jul, 2012 08:41:11 Message: 19 of 24 
"Sargondjani" wrote in message <juf082$gcm$1@newscl01ah.mathworks.com>... 
Subject: Scaling problem with fmincon From: Sargondjani Date: 22 Jul, 2012 10:33:30 Message: 20 of 24 
Thanks both. From your comments and some more tests, i am convinced that getting the finite differences right is much more important than getting the scaling right. 
"Sargondjani" wrote in message <jugktq$aot$1@newscl01ah.mathworks.com>... 
Subject: Scaling problem with fmincon From: Sargondjani Date: 24 Jul, 2012 06:12:09 Message: 22 of 24 
> > 4) in your 'pet example' the scaling did improve convergence (as reflected in the condition number), so i was expecting some similar improvement if i scaled my problem, 
"Sargondjani" wrote in message <julebp$e6j$1@newscl01ah.mathworks.com>... 
"Sargondjani" wrote in message <julebp$e6j$1@newscl01ah.mathworks.com>... 
No tags are associated with this thread.
A tag is like a keyword or category label associated with each thread. Tags make it easier for you to find threads of interest.
Anyone can tag a thread. Tags are public and visible to everyone.