Improving generalization of neural SAT solvers through algorithmic reasoning

Combinatorial optimization problems present a significant challenge due to their inherent complexity, naturally exponential, making algorithms impractical. Of special interest is the Boolean satisfiability problem (SAT) due to its wide-ranging applications in software and hardware verification, arti...

Descripción completa

Detalles Bibliográficos
Autor: Aranda Llorens, Oriol
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/410881
Acceso en línea:https://hdl.handle.net/2117/410881
Access Level:acceso abierto
Palabra clave:Combinatorial optimization
Neural networks (Computer science)
Machine learning
Deep learning (Machine learning)
Raonament Algorítmic Neuronal
SAT Solvers Neuronals
Problema de Satisfacibilitat Booleana
Xarxes Neuronals per Grafs
Xarxes Neuronals
Aprenentatge de Representacions per Grafs
Aprenentatge Profund
Aprenentatge Automàtic
Intel·ligència Artificial
Optimització Combinatòria
DPLL
Algoritmes
Grafs
Neural Algorithmic Reasoning
Neural SAT solvers
Graph Neural Networks
Graph Representation Learning
Neural Networks
Deep Learning
Machine Learning
Artificial Intelligence
SAT solving
Boolean Satisfiability Problem
Algorithms
Combinatorial Optimization
Graphs
Optimització combinatòria
Xarxes neuronals (Informàtica)
Aprenentatge automàtic
Aprenentatge profund
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic
Descripción
Sumario:Combinatorial optimization problems present a significant challenge due to their inherent complexity, naturally exponential, making algorithms impractical. Of special interest is the Boolean satisfiability problem (SAT) due to its wide-ranging applications in software and hardware verification, artificial intelligence, cryptography and optimization. Leveraging the significant advances in deep learning, specially Graph Neural Networks (GNNs), neural combinatorial SAT solvers have emerged as a promising alternative for solving SAT. However, their limited generalization and extrapolation capabilities make them less competitive than classical solvers. In recent years, Neural Algorithmic Reasoning has gained traction by altering the conventional approach of how machines learn to solve tasks. Rather than just learning to map inputs to outputs, this method focuses on understanding and replicating the reasoning steps of an algorithm, inherently characterized by a strong generalization. In this work, we introduce novel GNN-based models that leverage neural algorithmic reasoning, and we empirically show their great potential to enhance their ability to extrapolate. This exploration offers a fresh perspective on how the synergy between deep neural networks and classic algorithmic reasoning, can benefit neural combinatorial optimization and SAT solving.