The research introduces a neuro-symbolic framework designed to automate the process of constraint acquisition, a task traditionally reliant on extensive human interaction. Existing methods frequently require significant time and numerous queries to a human oracle. This new approach utilizes a neural Oracle Transformer to emulate user responses and generalize conceptual knowledge. The transformer interacts with a dedicated constraint acquisition engine, FastCA, which systematically refines the oracle’s responses into a sound, consistent, and interpretable constraint network. This interaction aims to align data-driven pattern recognition with symbolic reasoning. The framework is trained on previously available examples. The research demonstrates the effectiveness of this neuro-symbolic interplay in constructing robust models for combinatorial domains. Source: https://arxiv.org/abs/2609.12267
Loading
