document CCSA

Ignore-this: 81c81ee37e294f64a00358c52c3cf883

darcs-hash:20120720202035-c8de0-349519e0f396880f305dee17f6490d609bffbd3c.gz

diff --git a/mma/README b/mma/README
index 01c57c37ffd6065c79cab586a0cdae09bdd2d157..cb87086c2eccd9ad19cb4e5323e299063aab4a43 100644 (file)
--- a/mma/README
+++ b/mma/README
@@ -11,8 +11,22 @@ algorithms in NLopt, in that it handles an arbitrary set of nonlinear
 I've implemented the full nonlinear-constrained MMA algorithm, and it
 is exported under the nlopt_minimize_constrained API.
 
+I also implemented another CCSA algorithm from the same paper: instead of
+constructing local MMA approximations, it constructs simple quadratic
+approximations (or rather, affine approximations plus a quadratic penalty
+term to stay conservative).  This is the ccsa_quadratic code.  It seems
+to have similar convergence rates to MMA for most problems, which is not
+surprising as they are both essentially similar.  However, for the quadratic
+variant I implemented the possibility of preconditioning: including a
+user-supplied Hessian approximation in the local model.  It is easy to
+incorporate this into the proof in Svanberg's paper, and to show that
+global convergence is still guaranteed as long as the user's "Hessian"
+is positive semidefinite, and it practice it can greatly improve convergence
+if the preconditioner is a good approximation for (at least for the
+largest eigenvectors) the real Hessian.
+
 It is under the same MIT license as the rest of my code in NLopt (see
 ../COPYRIGHT).
 
 Steven G. Johnson
-July 2008
+July 2008 - July 2012