On deriving efficient information-seeking behaviour for intelligent autonomous systems

Abstract

Autonomous systems must learn and adapt efficiently in uncertain, non-stationary environments. This talk presents a principled mathematical framework for deriving information-seeking behavior in intelligent agents, grounded in Bayesian inference and the free energy principle. We formalize the problem of autonomous learning as variational free energy minimization, to be used both for inferring the parameters of a generative model and for inferring control policies. As an example of this approach, we introduce the ARxI agent, an autoregressive reactive inference system that leverages Bayesian message passing on factor graphs. By unrolling the generative model into the future and intervening on predictive distributions, ARxI optimizes a trade-off between information gain and goal-directed actions. This approach yields a variational posterior over controls that is analytically tractable and interpretable, providing insights into the role of uncertainty in adaptive decision-making. Finally, the talk will also highlight RxInfer.jl (https://rxinfer.com), an open-source toolbox for automatic real-time Bayesian inference, developed at TU Eindhoven.