Allen Downey: Bayesian Decision Analysis [Tutorial] | PyData Global 2022

Learn Bayesian decision analysis with PyMC, a powerful framework for estimation and inference. Update your beliefs using data, and don't rely solely on prior knowledge.

Key takeaways

Bayesian decision analysis involves updating beliefs based on data.
The speaker used a simple problem to demonstrate Bayesian decision analysis: a multi-armed bandit with four machines, each with a different probability of winning.
The prior distribution represents the initial beliefs about the machines.
The likelihood function represents the probability of observing the data given the hypothesis.
The posterior distribution represents the updated beliefs after observing the data.
The speaker used PyMC to estimate the posterior distribution.
Thompson sampling is a strategy for choosing a machine to play based on the posterior distribution.
Frequentist and Bayesian statistics can be used for the same problem, but Bayesian statistics provide a more informative prior and posterior distribution.
The speaker emphasized the importance of updating beliefs based on data and not relying solely on prior knowledge.
The prior distribution should be informed by domain knowledge, but not dominated by it.
The likelihood function can be estimated or modeled based on the data.
The posterior distribution provides a range of possible values for the parameter of interest, rather than a single point estimate.
Thompson sampling can be used to choose a machine to play based on the posterior distribution.
The speaker recommended using a uniform prior distribution as a default, but emphasized that it may not always be appropriate.
The speaker also recommended using PyMC for Bayesian inference, as it provides a convenient and efficient way to estimate posterior distributions.
The speaker’s final recommendation was to practice using Bayesian decision analysis with real-world problems.

Allen Downey: Bayesian Decision Analysis [Tutorial] | PyData Global 2022

More talks