ANR PRME Ceradoc

Creating Algorithms for Compact Decomposable Circuits

As artificial intelligence (AI) systems become increasingly integrated into sensitive domains, the need for transparency, reliability, and explainability becomes paramount. This project aims to enhance trust in AI by leveraging knowledge compilation techniques, particularly through the optimization of d-DNNF (Deterministic Decomposable Negation Normal Form) representations. By examining the fundamental properties of d-DNNF, we will focus on developing methods to create more compact circuit representations while preserving the computational capabilities of the language.

More specifically, the project will concentrate on developing heuristics for semantic decomposition, optimizing the sharing of sub-circuits, and enriching our understanding of various functions suitable for weighted model counting. Additionally, we will explore the potential of mixed representations that combine d-DNNF with other representation languages to facilitate model counting certification and reduce computation times when querying these circuits. Through thorough theoretical analysis and practical implementation, our work aims to design tools that can certify AI systems, thereby contributing to their transparency and reliability in critical applications such as neuro-symbolic AI and probabilistic modeling.