Bayesian Interest Group Virtual Meeting
Date: April 22, 2026
Time: 8:00am PT
Location: Via Zoom Meeting (RSVP to get the Zoom Link)
Putting into practice a principled Bayesian approach to decision-making demands a novel admixture of techniques from statistics, AI, and the fundamentals of applied decision theory. The pieces have been lying around; it just takes some insight to put them together. In this talk, I’ll walk you through an example that draws together these techniques in a workable framework.
We start by eliciting causal influences to form a network of variables linked by their conditional probabilities. We term this Bayes Network a “structural prior.” The network probabilities can be derived equally from data, or from judgment — Bayesian inference associates judgment with priors and statistical artifacts with likelihoods but treats them the same. Conventional machine learning methods do not make the prior explicit, but directly estimate posterior predictive distributions. Explicit statistical estimation of likelihoods avoids this confusion.
We work through one example of how this can be achieved with a classification tree application as used in statistics, illustrating how current machine learning methods can be applied in a principled way to Decision Analysis.
To complete a decision model, we extend the inference network with decision and value variables to create a Decision Network, traditionally called an Influence Diagram. Not only is this representation isomorphic to conventional probability decision trees—with the added benefit of clearly revealing the dependency structure of the problem, but it admits to efficient, scalable, and mature computational solution algorithms.
The result is a novel analytical approach that combines Influence Diagram solvers with Bayesian statistical methods, advancing the computational foundations of Decision Analysis and fulfilling what we call the “Bayesian Promise”: the coherent, unified treatment of judgment and empirical evidence through the calculus of probability.
About the speaker:
John Mark Agosta is co-founder of Fondata, LLC, and serves as an adjunct instructor at Stanford (Decision Analysis) and San José State University (Engineering Extended Studies). He previously spent eight years at Microsoft as part of the Azure Data AI Co-innovation Engineering Team.
With over 30 years of experience in applied research in the Bay Area, dating back to the early days of AI, he has held roles at SRI International, Toyota ITC, Intel Research, and several startups. He has authored 30+ peer-reviewed publications and holds six patents.
He founded the Bayesian Applications Workshop in 2002, held alongside the Uncertainty in AI conference, and was awarded a Business Fellowship at the Santa Fe Institute in 2007. He holds a BS in Physics from Yale University, an MS in Science Policy from George Washington University, and a PhD in Management Science and Engineering from Stanford University.
Interested in joining the Bayesian Interest Group?
Click here to learn more and join the Bayesian Interest Group: https://www.decisionprofessionals.com/communities/community-home?CommunityKey=a66c0277-adc7-44e5-bd43-019b098b682b