Integrating Constraints via Probabilistic Circuits

Soroush Ghandi, Benjamin Quost, Cassio de Campos

Keywords: Probabilistic circuits, constraints, probabilistic graphical models

Abstract: One of the recent advances in the domain of Probabilistic Circuits (PCs) is the introduction of methodologies for incorporating constraints into the represented distributions, thereby enabling the integration of external sources of information. In this paper, we investigate the extension of such paradigms to other classes of probabilistic models. In particular, we consider four representative models: continuous mixtures of tractable probabilistic models, Bayesian networks, Chow-Liu trees, and decision trees. We show that principled extensions of the techniques developed for PCs can be effectively applied to these models, thereby facilitating constrained optimization within a broader class of probabilistic frameworks.

Submission Number: 17