Predictive Processing: Exploring Multivalent Variables
Our research extends the causal Bayesian network implementation of the Predictive Processing Theory to account for multivalent variables. We also propose a framework for solving the exploration-exploitation trade-o in the Bayesian Predictive Processing implementation. Here we use a Q-Learning approach with Dirichlet distributions as hyperpriors and the free-energy principle as a base for learning. The latter links the proposed methods to neural mechanisms in the brain which have been linked to the exploration/exploitation trade-o . We tested our methods via behavioural studies where a robot had to learn an environment from scratch to navigate to a light source.
Faculteit der Sociale Wetenschappen