From: stevenj <stevenj@alum.mit.edu>
Date: Mon, 5 Apr 2010 00:33:44 +0000 (-0400)
Subject: allow COBYLA to increase trust-region radius (rho) if predicted improvement was appro... 
X-Git-Url: http://www.chiark.greenend.org.uk/ucgi/~ianmdlvl/git?a=commitdiff_plain;h=3dd3526a0c76bfa049b0c2e59cf8ebc7d8452759;p=nlopt.git

allow COBYLA to increase trust-region radius (rho) if predicted improvement was approximately right and simplex is ok, following suggestion in SAS manual for PROC NLP

darcs-hash:20100405003344-c8de0-d288279ade2dfafe7b9bcf9facf7807288f28010.gz
---

diff --git a/cobyla/cobyla.c b/cobyla/cobyla.c
index c45b626..ea34c89 100644
--- a/cobyla/cobyla.c
+++ b/cobyla/cobyla.c
@@ -939,6 +939,17 @@ L440:
 /* Branch back for further iterations with the current RHO. */
 
   if (trured > 0. && trured >= prerem * .1) {
+    /* SGJ, 2010: following a suggestion in the SAS manual (which
+       mentions a similar modification to COBYLA, although they didn't
+       publish their source code), increase rho if predicted reduction
+       is sufficiently close to the actual reduction.  Otherwise,
+       COBLYA seems to easily get stuck making very small steps. 
+       Also require iflag != 0 (i.e., acceptable simplex), again
+       following SAS suggestion (otherwise I get convergence failure
+       in some cases.) */
+    if (trured >= prerem * 0.9 && trured <= prerem * 1.1 && iflag) {
+	 rho *= 2.0;
+    }
     goto L140;
   }
 L550: