Local maximum with mixture model

Hi everyone,

Does anyone know how Bayesian inference handle the issue of local maxima on mixture models?

Thanks.

https://mc-stan.org/users/documentation/case-studies/identifying_mixture_models.html

Is a good place.

Bayesian inference doesn’t ‘handle’ multimodal posteriors in mixtures. You have to identify the model such that it de-aliases the mixture components (i.e., reduces the multimodality down to one mode).

As for ‘local modes’, which are not the result of an unidentified model, the hope is that the sampling scheme will be able to traverse into various modes, and therefore effectively integrate over the lumpy posterior surface. I.e., hopefully the sampler doesn’t get stuck in a local mode, but just keeps sampling across them.

In either case, so long as you start multiple chains from multiple starting positions, you should have a decent idea of whether the sampler has found a common space via the Rhat diagnostic.

2 Likes

Hi Stephen,

Thank you for sharing the link. Your answer is very helpful to me.

Doria

1 Like