Hands-On Q-Learning with Python

Bayesian control rule

The Bayesian control rule is a generalization of the Thompson sampling process that has the agent develop general rules about causal relationships in the environment. Briefly, as it interacts with the environment, it codifies these causal relationships and adopts a behavior that maximizes its own utility with respect to the environment.

Here is a brief primer on Bayes' rule and priors: in probability terms, a prior distribution is our current estimate of what the likelihood of an event is, and a posterior distribution is our estimate of the likelihood of that event once we run a trial on it and get additional information about it.

This is the Bayesian approach to assigning probability distributions to unknown future ...