Overview
Propagate the multi-start gradient MAP searches (PyAutoFit#1369, shipped: af.MultiStartAdam / MultiStartADABelief / MultiStartLion) to this workspace's scripts/guides/modeling/searches.py guide, alongside the existing LBFGS section.
Plan
- Add a MultiStartAdam section to
scripts/guides/modeling/searches.py after the LBFGS section: a af.MultiStartAdam(...) config block + prose. The guide is a configuration reference (it constructs search objects, does not run fits), so this mirrors how LBFGS is shown.
- Frame it against the caveat the guide already states (optimizers struggle on these complex parameter spaces): multi-start Adam is the optimizer that works, because its wide population of parallel starts escapes the local maxima that trap single-start LBFGS (benchmark-proven on the HST MGE lens likelihood). Note it needs
use_jax=True; MultiStartADABelief/MultiStartLion are drop-in alternatives; it returns a MAP point estimate (Nautilus stays default for errors).
- Update the
__Contents__ list and regenerate the paired notebooks/guides/modeling/searches.ipynb.
Phase 3 of the multi-start gradient search promotion. --auto, effective level safe (docs cap).
Overview
Propagate the multi-start gradient MAP searches (PyAutoFit#1369, shipped:
af.MultiStartAdam/MultiStartADABelief/MultiStartLion) to this workspace'sscripts/guides/modeling/searches.pyguide, alongside the existing LBFGS section.Plan
scripts/guides/modeling/searches.pyafter the LBFGS section: aaf.MultiStartAdam(...)config block + prose. The guide is a configuration reference (it constructs search objects, does not run fits), so this mirrors how LBFGS is shown.use_jax=True;MultiStartADABelief/MultiStartLionare drop-in alternatives; it returns a MAP point estimate (Nautilus stays default for errors).__Contents__list and regenerate the pairednotebooks/guides/modeling/searches.ipynb.Phase 3 of the multi-start gradient search promotion.
--auto, effective level safe (docs cap).