configurable tolerance for gradient stopping condition (#121) #585
Conversation
src/api/optimize.c
Outdated
| @@ -556,6 +556,9 @@ static nlopt_result nlopt_optimize_(nlopt_opt opt, double *x, double *minf) | |||
| stop.ftol_abs = opt->ftol_abs; | |||
| stop.xtol_rel = opt->xtol_rel; | |||
| stop.xtol_abs = opt->xtol_abs; | |||
| stop.tolg = (opt->tolg > 0.) | |||
| ? opt->tolg | |||
| : 1e-8; /* SGJ: was 1e-6, but this sometimes stops too soon */ | |||
There was a problem hiding this comment.
lets use the get_param/set_param api for this instead of adding it for all algorithms
that way it will also be accessible by languages bindings
and add a documentation entry
There was a problem hiding this comment.
Indeed, that's what I was looking for!
Move 'tolg' to generic parameters
src/api/optimize.c
Outdated
| @@ -735,7 +735,7 @@ static nlopt_result nlopt_optimize_(nlopt_opt opt, double *x, double *minf) | |||
| case NLOPT_LD_TNEWTON_PRECOND: | |||
| case NLOPT_LD_TNEWTON_PRECOND_RESTART: | |||
| #ifdef NLOPT_LUKSAN | |||
| return luksan_pnet(ni, f, f_data, lb, ub, x, minf, &stop, opt->vector_storage, 1 + (algorithm - NLOPT_LD_TNEWTON) % 2, 1 + (algorithm - NLOPT_LD_TNEWTON) / 2); | |||
| return luksan_pnet(ni, f, f_data, lb, ub, x, minf, &stop, opt->vector_storage, 1 + (algorithm - NLOPT_LD_TNEWTON) % 2, 1 + (algorithm - NLOPT_LD_TNEWTON) / 2, opt->get_param("tolg", 0.)); | |||
There was a problem hiding this comment.
maybe we could set the default value here and simplify the rest instead of using 0 which only means use the default that will be set later
There was a problem hiding this comment.
I considered your idea, what stopped me from choosing it was :
- using a default here won't cover the case of the user entering an invalid value (negative or null is currently switched back to default)
- I find that the default value belongs more legitimately to the algorithm than to the generic caller
- it is technically cumbursome to define a variable inside a
case(you need braces)
None of these items is a real issue. Their sum oriented me toward my current proposal, but I totally agree if you prefer the other way.
There was a problem hiding this comment.
ok, but it doesnt build though
There was a problem hiding this comment.
(shame on me!) my apologies...
#121
Luksan algorithms (BFGS, troncated Newton, Shifted variable-metric) use a hard-coded value for their gradient stopping condition.
The tolerance is exposed to the API, with current value as default.