Multi-Start: Michaelis-Menten Elimination
Michaelis-Menten (saturable) elimination is the canonical example of a local-minimum problem in NLME modelling. The Vmax/Km pair is only weakly identifiable from sparse PK data: many (Vmax, Km) combinations produce similar predicted concentrations, so the OFV surface contains a narrow ridge. A single-start FOCEI run is sensitive to the initial values — starting far from the true optimum, the optimizer often settles on an inflated Vmax/Km pair with a higher OFV than the true optimum.
The problem
For the model below with true parameters Vmax ≈ 3.5 mg/h and Km ≈ 5.5 mg/L, starting with Vmax = 12 and Km = 20 (both roughly 3× too high) a single FOCEI run may converge to a local minimum rather than the global one. The two solutions predict similar concentration-time profiles on the observed time grid but differ in OFV by several units.
The model file
{{#include ../../../examples/mm_multistart.ferx}}
What multi-start does
With n_starts = 8 and start_sigma = 0.5:
- Start 0 uses the exact initial values from
[parameters](Vmax=12, Km=20). - Starts 1–7 multiply each log-packed theta by
exp(N(0, 0.5)), giving a spread of roughly ×0.6–×1.6 around the initial values. Withstart_sigma = 0.5this is wider than the default 0.3 — appropriate for a ridge-shaped surface. - All 8 runs execute in parallel. On an 8-core machine wall-clock time equals a single run.
- The converged run with the lowest OFV is returned. If start 0 was already at the global optimum, the estimates are those of a single run to within the optimizer’s stopping noise (the returned start can be another one that reached the same optimum).
- One
Multi-start:line onFitResult.warningssays how the starts compare, andFitResult.startshas a row per start.
Running it
# Simulate from the [simulation] block and fit in the same run
cargo run --release -p ferx-cli -- examples/mm_multistart.ferx --simulateReading the result
The run prints (macOS arm64; the winning index can differ by platform, since the starts agree to 1e-7):
Multi-start: 8 starts, 8 converged; best OFV -221.8770 (start 6); all 8 converged starts within 1.0 of best (range 0.0000)
All eight starts reached the same optimum. Start 6 is lower than start 0 by 1e-7 OFV; that is the optimizer’s stopping noise, not a better optimum. This is the answer multi-start is run for: the reported optimum does not depend on the initial values.
Widen the perturbation to start_sigma = 1.5 and the line changes:
Multi-start: 8 starts, 8 converged; best OFV -221.8770 (start 2); 7 of 8 converged starts within 1.0 of best; 1 converged above it: start 4 (+5.27)
Start 4 stopped 5.27 OFV above the others and reports converged. Its row in FitResult.starts (and in the fit YAML’s multi_start: block) shows where: TVVMAX at its upper bound of 50, with TVV 3.9 against 18.5. The returned estimates are the same as before. What the line adds is that this objective has a second point where a fit can stop, so a single run from a poor start could have reported it. What the starts did explains each part of the line.
When to use multi-start
| Model feature | Risk of local minima | Recommended n_starts |
|---|---|---|
| Linear 1-cpt / 2-cpt | Very low | 1 (default) |
| Michaelis-Menten elimination | High | 4–8 |
| Full-block omega (≥ 3 etas) | Moderate | 4 |
| Many correlated covariates | Moderate | 4–8 |
| Small dataset (e.g. an arm-level MBMA) | Model-dependent; cheap to check | 8 |
Small datasets
On a dataset of a few hundred rows or fewer, such as an arm-level model-based meta-analysis, a single fit usually takes well under a second, so eight starts cost almost nothing. That makes multi-start a cheap check that the reported optimum is not a local one, even when the model is not one of the high-risk types above. The exception is a model whose nested levels stack many random effects into one subject: on 192 rows such a fit took 0.56 to 6.97 s as the stacked width grew from 9 to 33 (Cost grows with the groups inside each subject).
Whether it finds anything depends on the model. On the naproxen MBMA (examples/mbma_naproxen.ferx, 18 trials, 36 arms, 122 observations) it found nothing better: from the published estimates and from a naive start (all covariate effects at 0, the typical values off by 2–7×), n_starts = 1 and n_starts = 8 reached the same OFV, 261.8149, and eight starts took 0.04–0.06 s against 0.03 s for one. A user fitting arm-level ulcerative-colitis models reported better optima from multi-start on several of them (ferx-r#721). The Multi-start: line says which it is: all 8 converged starts within 1.0 of best means the single-start fit was already there, whichever start index the line names, and a clause such as 1 converged above it: start 4 (+5.27) means the starts disagree (Reading the result).
start_sigma = 0.3 (the default, ≈ 30% CV) is appropriate for most models. For ridge-shaped surfaces like Vmax/Km, start_sigma = 0.4–0.5 explores a wider neighbourhood.