```Date: Mon, 30 Apr 2001 12:07:59 -0700 Reply-To: Dale McLerran Sender: "SAS(r) Discussion" From: Dale McLerran Subject: Re: How to minimize step functions? Comments: To: jweedon@earthlink.net Content-Type: text/plain Jay, The first place I start when trying to resolve difficulties with nonlinear optimization problems is to generate and plot the function value for several widely spaced points in space. This does a couple of things for you: 1) it helps you to visualize the basic shape of the response function which can give you some clues as to what initial values work best and whether the surface is poorly defined, and 2) it can help identify programming errors. If the response is very flat, it may point you to some term which you misspecified. OK, it probably doesn't happen to you that the original function code gets written incorrectly, but it does happen to me. Good luck, Dale >Date: Mon, 30 Apr 2001 09:26:59 -0400 >Reply-To: Jay Weedon >From: Jay Weedon >Subject: How to minimize step functions? >To: SAS-L@LISTSERV.UGA.EDU > >Hi folks, > >I've been having trouble getting the IML nonlinear optimization >routines to solve a particular minimization problem. They converge >after one iteration, i.e., never improve on my starting value. > >I suspect that the problem is that the function, though continuous, is >locally flat everywhere, i.e., is a step function. The number of steps >is unknown (though probably runs to dozens or possibly a few hundred), >and I do not know the locations of the step boundaries. Nor do I know >whether there exist local as well as global minima. I do know the >boundaries of the numerical range within which the global minimum must >exist. I also know that the steps cannot be further apart than some >specific small value (e.g., 0.0001), which I suppose provides an upper >limit to the number of steps. > >I'm not mathematically sophisticated enough to know how to deal with >this situation as a general numerical problem (other than doing a >binary search). Are there SAS routines that can handle this? Or can >someone point me to a numerical techniques recipe so I can write my >own code? > >TIA, >Jay Weedon. --------------------------------------- Dale McLerran Fred Hutchinson Cancer Research Center mailto: dmclerra@fhcrc.org Ph: (206) 667-2926 Fax: (206) 667-5977 --------------------------------------- ------------------------------------------------------------ --== Sent via Deja.com ==-- http://www.deja.com/ ```

Back to: Top of message | Previous page | Main SAS-L page