跳转至

文章背景与核心概要

本文探讨了逻辑张量网络增强型生成对抗网络（LTN-GANs）中如何处理硬约束（如排序、正定性以及定义恒等式）这一关键技术难题。传统的 LTN-GANs 通常将逻辑公理基础化为谓词（predicate），通过优化模糊真值来引导生成器，但这仅能对样本进行打分，无法从结构上强制满足严格的硬约束。

为了解决这一局限性，作者研究了在 LTN 框架内将每个公理基础化为“函数符号（function symbol）”的方法，并将其与现有的边界裁剪（boundary-clamping）替代方案进行了对比。研究发现，数学上有效（Valid）的样本并不等同于现实中逼真的样本，不等式成立时往往带有余量（margin），而高保真的生成器必须捕捉到这些余量的实际分布。作者引入了分辨率比（\(R\)）这一预训练诊断指标，揭示了传统方法在处理高 \(R\) 值时会导致余量分布丢失的缺陷。而函数符号通过直接计算受约束的变量，在可行域内部建立了一个坐标系（即图表，chart），从而确保每个样本在结构上天然有效，同时完美学习到底层的余量分布。

Generate in the Chart, Not on the Boundary: Function-Symbol Grounding for Hard Constraints in LTN-GANs

Generate in the Chart, Not on the Boundary: Function-Symbol Grounding for Hard Constraints in LTN-GANs

Authors: Nijesh Upreti, Vaishak Belle
Primary Subject: Artificial Intelligence (cs.AI)
arXiv ID: 2608.21605
Submitted: 21 August 2026

📌 Summary

📌 Summary

逻辑张量网络增强型生成对抗网络（LTN-GANs）通过将逻辑公理基础化为谓词（predicates），将背景知识注入模型中，从而允许生成器朝著更高的模糊真值（\(0\) 到 \(1\) 之间的满足度分数）进行优化。

Logic Tensor Network-Enhanced Generative Adversarial Networks (LTN-GANs) inject background knowledge into models by grounding logical axioms as predicates, allowing the generator to optimize toward higher fuzzy truth values (satisfaction scores between \([0,1]\)).

虽然传统的谓词级基础化有助于整体上满足约束，但它仅能对样本进行打分，而无法嵌入那些必须在每个生成样本中严格成立的硬结构约束（例如排序、正定性和定义恒等式）。

While traditional predicate-level grounding helps satisfy constraints overall, it only scores samples and falls short of embedding hard structural constraints (like orderings, positivity, and definitional identities) that must strictly hold for every generated sample.

本文研究了在 LTN 框架内将每个公理基础化为函数符号（function symbol）而非谓词的方法，并将其与边界裁剪（boundary-clamping）这一替代方法进行了对比。核心洞察包括： * 有效性谬误（The Validity Fallacy）： 通过边界裁剪在数学上有效的样本并不一定逼真。不等式是以一定的余量（margin）成立的，一个高保真的生成器必须捕捉到这些余量的实际分布。 * 分辨率比（\(R\)）： 作者引入了 \(R\)（数据尺度相对于余量扩展的比值）作为一项预训练诊断指标，以确定给定的基础化可以学习哪些约束。高 \(R\) 值会导致谓词方法在标准指标看起来不错的同时静默失效，因为裁剪会将每个样本推向边界并抹去余量分布。 * 函数符号基础化（Function-Symbol Grounding）： 通过直接计算受约束的变量而非仅仅对其打分，函数符号建立了一个图表（chart）——即可行域内部的一个坐标系。这确保了每个样本在构造上就是有效的，同时成功地像学习其他量一样学习了余量分布。

This paper investigates grounding each axiom as a function symbol within the LTN framework rather than a predicate, comparing it against the alternative boundary-clamping method. The key insights include: * The Validity Fallacy: A sample that is mathematically valid (via boundary clamping) is not necessarily realistic. Inequalities hold by a margin, and a faithful generator must capture the actual distribution of those margins. * The Resolution Ratio (\(R\)): The authors introduce \(R\) (data scale relative to margin spread) as a pre-training diagnostic to determine which constraints a given grounding can learn. High \(R\) values cause predicate methods to fail silently while standard metrics look fine, as clamping pushes every sample onto the boundary and erases margin distributions. * Function-Symbol Grounding: By computing the constrained variable directly instead of merely scoring it, function symbols establish a chart—a coordinate system inside the feasible region. This ensures every sample is valid by construction while successfully learning the underlying margin distribution.

📑 Abstract

📑 Abstract

Logic Tensor Network-Enhanced Generative Adversarial Networks (LTN-GANs) inject background knowledge by grounding each logical axiom as a predicate and training the generator to raise its satisfaction, a fuzzy truth value in \([0,1]\). Previous LTN-GAN work grounded every constraint this way, at the predicate level, and improved constraint satisfaction. A predicate, however, only scores a sample, so it cannot embed hard structural constraints, rules such as orderings, positivity, and definitional identities that must hold in every generated sample. In this work, we investigate grounding each axiom as a function symbol inside the LTN framework. We compare against the state-of-the-art alternative, a constraint layer that clamps each violating sample onto the feasible boundary and so produces outputs that are always valid. Our investigation shows that a valid sample is not always a realistic one. An inequality is not merely satisfied or violated. It holds by a margin, and a faithful generator should also reproduce the margin's real distribution. We find that the resolution ratio \(R\), the data's scale over the margin's spread, is a diagnostic, computable before training, of which constraints a chosen grounding can learn. When \(R\) is large, the predicate receives no learning signal, the clamp pushes every sample onto the boundary, and the margin distribution is lost while every standard metric still looks fine. A function symbol avoids both failures, computing the constrained variable rather than scoring it. Together the function symbols form a chart, a coordinate system inside the feasible region, where every sample is valid by construction and the margin is learned like any other quantity.