Symbolic Artificial Intelligence has been considered ''Good Old-Fashioned Artificial Intelligence'' since it usually represents knowledge through explicit symbols, such as first-order logic predicates and constants, instead of through large numeric matrices, as happens for sub-symbolic solutions, namely, neural networks. Despite the recent predominance of neural networks, considered almost a silver bullet for solving every machine learning problem, there is still an ever-growing field of research called Statistical Relational Artificial Intelligence, where the goal is to combine logic and uncertainty to represent and reason over complex domains. Among all the several possible formalisms, Probabilistic Logic Programming (PLP) is gaining traction thanks to its ability to integrate logical and probabilistic reasoning. The work presented in this dissertation is structured in two parts. In the first, we present several extensions of PLP to widen the possible application scenarios. In particular, we review hybrid probabilistic logic programs, where continuous and discrete random variables coexist, and provide a new well-defined semantics. After this, we introduce probabilistic abductive logic programs, where we extend PLP with the possibility to reason with incomplete data, and probabilistic optimizable and reducible logic programs, where we leverage constraint programming to learn the parameters and the structure of probabilistic logic programs subject to constraints. The definition of these new classes is motivated by real-world examples discussed in the correspondent chapters. For all of these proposals, we formally introduce the task to solve and provide practical inference algorithms. The second part is focused on the adoption of PLP and hybrid probabilistic logic programs to model several blockchain-related scenarios. We discuss models to analyse transaction fees, smart contracts, double spending, and the Lightning Network.
L'intelligenza artificiale simbolica viene definita come ''Good Old-Fashioned Artificial Intelligence'' perché rappresenta ciò che è noto tramite simboli (per esempio, predicati del primo ordine e costanti) invece che tramite matrici di numeri, come accade per le soluzioni sub-simboliche (principalmente reti neurali). Nonostante il recente predominio delle reti neurali, considerate quasi una panacea per risolvere qualsiasi problema legato all'apprendimento automatico, esiste un ambito di ricerca in rapida espansione chiamato Statistical Relational Artificial Intelligence dove l'obiettivo è quello di combinare ragionamento logico ed incertezza per rappresentare ed estrarre conoscenza da domini complessi. Tra i possibili formalismi, la Programmazione Logico Probabilistica (PLP) sta suscitando molto interesse grazie alla possibilità di integrare logica e probabilità. Il lavoro presentato in questa dissertazione è strutturato in due parti. Nella prima parte vengono analizzate diverse estensioni per PLP che ampliano i possibili scenari applicativi. In particolare, vengono analizzati programmi logico probabilistici ibridi che presentano sia variabili continue che discrete. Per questi, viene fornita una nuova semantica ben definita. Successivamente, vengono introdotti programmi logico probabilistici abduttivi, che estendono PLP con la possibilità di gestire dati incompleti, ottimizzabili e riducibili, dove viene utilizzata la programmazione a vincoli per apprendere i parametri e la struttura di programmi logico probabilistici con vincoli. L'introduzione di queste nuove classi è motivata da diverse possibili applicazioni discusse nelle corrispondenti sezioni. Per ciascuna nuova estensione, viene definito formalmente il problema che si vuole risolvere e vengono forniti algoritmi per effettuare inferenza. La seconda parte è incentrata sulla rappresentazione tramite PLP e programmi ibridi di scenari tipici di sistemi blockchain. Vengono discussi modelli per analizzare le commissioni associate alle transazioni, smart contract, double spending e la rete Lightning Network.
Extensions and Applications of Probabilistic Logic Programming
AZZOLINI, DAMIANO
2022
Abstract
Symbolic Artificial Intelligence has been considered ''Good Old-Fashioned Artificial Intelligence'' since it usually represents knowledge through explicit symbols, such as first-order logic predicates and constants, instead of through large numeric matrices, as happens for sub-symbolic solutions, namely, neural networks. Despite the recent predominance of neural networks, considered almost a silver bullet for solving every machine learning problem, there is still an ever-growing field of research called Statistical Relational Artificial Intelligence, where the goal is to combine logic and uncertainty to represent and reason over complex domains. Among all the several possible formalisms, Probabilistic Logic Programming (PLP) is gaining traction thanks to its ability to integrate logical and probabilistic reasoning. The work presented in this dissertation is structured in two parts. In the first, we present several extensions of PLP to widen the possible application scenarios. In particular, we review hybrid probabilistic logic programs, where continuous and discrete random variables coexist, and provide a new well-defined semantics. After this, we introduce probabilistic abductive logic programs, where we extend PLP with the possibility to reason with incomplete data, and probabilistic optimizable and reducible logic programs, where we leverage constraint programming to learn the parameters and the structure of probabilistic logic programs subject to constraints. The definition of these new classes is motivated by real-world examples discussed in the correspondent chapters. For all of these proposals, we formally introduce the task to solve and provide practical inference algorithms. The second part is focused on the adoption of PLP and hybrid probabilistic logic programs to model several blockchain-related scenarios. We discuss models to analyse transaction fees, smart contracts, double spending, and the Lightning Network.File | Dimensione | Formato | |
---|---|---|---|
tesi_no_abstract.pdf
accesso aperto
Descrizione: tesi_no_abstract
Tipologia:
Tesi di dottorato
Dimensione
1.58 MB
Formato
Adobe PDF
|
1.58 MB | Adobe PDF | Visualizza/Apri |
I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.