PCiunkiewicz / dynamic-discretization

MIT License
1 stars 1 forks source link

Full DD algorithm #1

Open keycaliber-maggie opened 2 years ago

keycaliber-maggie commented 2 years ago

I've found this codebase quite useful for understanding many aspects of the dynamic discretization algorithm. However, I'm still not sure how to code the full algorithm given in Norman Fenton's Risk Assessment and Decision Analysis with Bayesian Networks, Second Edition, appendix D, p. 598. In this particular case no dataset is involved as in your implementation.

My question is, given a posterior marginal distribution for a node under a particular discretization, how do you calculate the bound on the KL distance as required for the next step? In particular, where do you find f_min, f_max, and f_bar, given that each interval only is assigned a single probability value as far as I can tell?

Sorry I'm putting this as an "issue," when it's more of a query. I couldn't find another way to message you. Thanks.

PCiunkiewicz commented 2 years ago

Hi @keycaliber-maggie, this problem becomes simple when we consider each interval a piece-wise uniform distribution. We can then approximate these values based on that assumption and compute the KL divergence. If you email me at philip.ciunkiewicz@ucalgary.ca, I can provide you with a pre-production copy of our accepted paper with further details on this methodology! Once the article is published and has a DOI, I will include it in this GitHub repository as I have done with other projects.

keycaliber-maggie commented 2 years ago

Thank you! I will email you.