New tutorial on optimal decisions using Utility Theory

TL;DR, here is the tutorial!

There is a wave towards using data more and more in decision making across all levels of business and society. However, people often use data quite informally: look at an Excel sheet or a graph, then make a decision based on your impression. This often works well, but it can be fragile because of our many deep-skin biases as well as a general poor ability to reason about quantities and complex interactions.

Decision theory to the rescue! By adding a few axioms to the basic axioms of probility theory, we can extend statistical modeling to make decisions which maximize utility – whether that utility is happiness, profit, health, public support, or something else entirely. I’m not saying that we should blindly hand over decisions to algorithms, but seeing their limited-worldview pure-quantitative solution can be a nice decision support to keep some of our biases at bay.

I was motivated to write this tutorial to fill in a gap: there is a lack of practical entry-level guides that scale well to complex problems. The entry-level here is someone who just want to update an Excel sheet with new data, and see how that changes the decision. Here’s the accompanying twitter thread:

I made a first draft of this for an elective course in the fall. Then a second draft for my presentation at the Bayes@Lund 2019 conference. And now I finally got to brush it off. OK, I become overly excited when I get to talk about Bayesian inference AND utility at the same time:

Jonas Kristoffer Lindeløv: extending bayes to make optimal decisions