Re: [Gretl-devel] testcases and pet projects

On Mon, 6 Jul 2009, Riccardo (Jack) Lucchetti wrote: Ok, I think I've got it. The reason why line 548 in gretl_bgfs.c reads crit_ok = !na(f) && (f >= fmax + d); instead of crit_ok = !na(f) && (f >= fmax); is that, in normal circumstances, we want to avoid jumping from one place to another if the jump doesn't yield a *significant* improvement in the objective function. Here d is the threshold, which, in turn, is computed as d = sumgrad * steplen * acctol; acctol is a hand-tweaked small number, steplen is between 0 and 1, so what really makes the difference here is sumgrad. The sumgrad variable contains the result of a quadratic form g'Hg, where g is the gradient and H is the BFGS curvature matrix. Now, at the first iteration this is conventionally set at I, while it could be any other pd matrix. So if for some problem the identity matrix happens to be a poor choice for H, then d becomes unreasonably big and you never have the chance of satisfying the criterion. This is not a problem after the first iteration, since H should contain "sensible" numbers, but the initial value for H becomes a problem. The following one-liner sets d=0 for the first iteration only: --- lib/src/gretl_bfgs.c.~1.7.~ 2009-07-06 11:07:30.000000000 +0200 +++ lib/src/gretl_bfgs.c 2009-07-06 12:06:48.000000000 +0200 @@ -543,7 +543,7 @@ } if (ndelta > 0) { f = cfunc(b, data); - d = sumgrad * steplen * acctol; + d = (iter>1) ? sumgrad * steplen * acctol : 0; fcount++; crit_ok = !na(f) && (f >= fmax + d); #if BFGS_DEBUG which lets your example script run fine. Could you please try it with your "real" stuff? Riccardo (Jack) Lucchetti Dipartimento di Economia Università Politecnica delle Marche r.lucchetti(a)univpm.it http://www.econ.univpm.it/lucchetti