# Using output from optimization algorithms to initialize sampler

**URL:** <https://discourse.mc-stan.org/t/using-output-from-optimization-algorithms-to-initialize-sampler/6969>\
**Category:** Algorithms\
**Tags:** optimization, mcmc\
**Created:** [December 14, 2018, 11:13am UTC](https://discourse.mc-stan.org/t/using-output-from-optimization-algorithms-to-initialize-sampler/6969 "2018-12-14T11:13:16Z")\
**Posts on this page:** 1\
**Showing post:** 7

<div class="post-metadata">

**Author:** ![monnahc](https://yyz2.discourse-cdn.com/flex030/user_avatar/discourse.mc-stan.org/monnahc/32/1936_2.png) [@monnahc](https://discourse.mc-stan.org/u/monnahc)\
**Post date:** [April 25, 2019, 6:51pm UTC](https://discourse.mc-stan.org/t/using-output-from-optimization-algorithms-to-initialize-sampler/6969/7 "2019-04-25T18:51:03Z")

</div>

I’m a bit late to this but you might want to browse this thread where I proposed the same idea. Some good discussion:

> [@Overdispersed initial values -- general questions](https://discourse.mc-stan.org/t/overdispersed-initial-values-general-questions/3966):
>
> I know the general advice is to initialize Stan (or any MCMC algorithm) from widely dispersed values relative to the posterior. My understanding is this helps catch issues like multimodality or chains not mixing well, which might not occur if started from the same point. But if the idea is to find local minima, why not use the optimizer? Start from the same dispersed values and optimize and check to see they all find the same mode (assuming for now it exists). It seems much more computationally…

---

_[View the full topic](https://discourse.mc-stan.org/t/using-output-from-optimization-algorithms-to-initialize-sampler/6969)._