Long CNL runtime causes

[Mohamed Khalifa]

Hello Prof. Bierlaire,

I was testing the runtime of different models in Apollo and Biogeme.  In Mixed Logit Models Biogeme is faster than Apollo , in Nested Logit Models Apollo is faster. It caught my attention that the difference in CNL models is too big. I tried investigating this and Biogeme uses more memory than in Apollo, but this is also the case for the other models as this is very common in python. I was wondering if you know what else can one investigate to find causes for this.

Best regards,

Mohamed Khalifa

[Michel Bierlaire]

The new version of Biogeme is based on Jax, that constructs graphs to calculate derivatives using backward differentiation. This significantly speeds up the estimation of mixture models, and also Bayesian estimation. It also better exploits parallelization on servers.
But it is true that it generates huge data structure for complex models like nested and cross-nested logit. I did not find the time to investigate it yet.

> --
> 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/35ce126c-a831-473c-94e6-c91c3e4b0450n%40googlegroups.com.
> <runtime_comparison.png>

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