Large Parameters estimation in Latent Class Choice Model

[Sourav Kr Mandal]

Respected Sir,

Is it possible to estimate 120-130 beta parameters for a 4 Class Latent Choice Modelling with N = 1600 using biogeme? Also, is there any chance that because of the complicated log-likelihood function, it might get optimised in local optima rather than in global optima?

[Michel Bierlaire]

> On 15 Jul 2025, at 20:39, Sourav Kr Mandal <mandals...@gmail.com> wrote:
>
> Respected Sir,
>
> Is it possible to estimate 120-130 beta parameters for a 4 Class Latent Choice Modelling with N = 1600 using biogeme? Also, is there any chance that because of the complicated log-likelihood function, it might get optimised in local optima rather than in global optima?

Certainly — there is a very high chance, even with fewer parameters. Latent class models tend to exhibit many local optima.

A student of mine recently developed a heuristic to help mitigate this issue (although without formal guarantees):
https://transp-or.epfl.ch/documents/technicalReports/HaerBier2025.pdf

We also proposed another heuristic some time ago, though it was never implemented in Biogeme:
https://dx.doi.org/10.1287/ijoc.1090.0343


>
> --
> You received this message because you are subscribed to the Google Groups "Biogeme" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to biogeme+u...@googlegroups.com.
> To view this discussion visit https://groups.google.com/d/msgid/biogeme/8533c517-eec9-4ee6-b355-58a6177e94f5n%40googlegroups.com.

Michel Bierlaire
Transport and Mobility Laboratory
School of Architecture, Civil and Environmental Engineering
EPFL - Ecole Polytechnique Fédérale de Lausanne
http://transp-or.epfl.ch
http://people.epfl.ch/michel.bierlaire