Hello,
I’ve been generous “book budget” from my employer to help build a “digital twin” of our supply chain so we can run simulations for “what-if” scenarios. Is this a problem that bayesian networks can solve? Does anyone have book recommendations for this?
Yes, you can use Bayesian Networks to run “what-if” scenarios provided your network can be interpreted causally. This means that nodes are not just linked because they seem to be related observationally, but because there’s a cause-effect connection between them.
For an introduction to Causal Inference (the discipline of discerning observational from causal relationships) you can read Judea Pearl’s: “The Book of Why: The New Science of Cause and Effect”. Incidentally Judea Pearl was the one who coined the term “Bayesian Networks” and devotes a chapter of the book to discussing them.
Thank you. They are definitely linked as the nodes would be supply chain related (manufacturing plant to distribution warehouse to store to customer).
I read the book of why two years ago at the urging of @JessieGrabowski but didn’t have a need to perform causal modeling at the time. I’ll brush back up.
Thanks again.
Bayesian networks is just another way of saying directed graphical model, which is just a specific way of specifying a generative joint probability distribution.
To get a feeling for the larger field if not for directed graphical models specifically, I’d recommend Richard McElreath’s Statistical Rethinking. It’s more about methodology for how to test model fits, how to do prediction, etc. There’s also Gelman et al.'s “Bayesian Worfklow” (arXiv), but in the sake of full disclosure, I’m a co-author on that one—it’s being expanded into a book of case studies. It’s largely about how to do applied Bayesian modeling and predictive inference. Both of these are based on Stan, not PyMC, but it’s a prettys simple translation between them for most problems.
Thank you @bob-carpenter. Yes, Statistical Rethinking was instrumental, especially when trying to re-think how we viewed demand forecasting. I’ll checkout the Bayesian Workflow paper as well. Thank you for pointing that out.