Augusto Pedron
A Fast Bayesian Artificial Intelligence Reasoning Engine For Modeling And Optimization Tasks.
Rel. Alessandro Savino, Stefano Di Carlo. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2022

PDF (Tesi_di_laurea)
 Tesi
Licenza: Creative Commons Attribution Noncommercial No Derivatives. Download (2MB)  Preview 
Abstract: 
A Bayesian Network is a graphical model used to visually represent connections between variables. The main task performed on a Bayesian Network is the inference of a posterior probability distribution of one or more nodes which represents the probability that each variable’s state has to occur. This probability distribution is used as a support to take decisions in the real world case which the network models. A factor to consider when taking decisions based on a posterior probability distribution is the influence that a parent node has on the node of interest. Based on the strength of influence measured, the user can alter its decision in accordance to the real world case. Before the network can become an useful tool that can assist the user, its structure and the prior probability distribution of each node has to be defined. Both the structure and the probability distributions can be either defined by hand or learnt through an input data set, making the modelling phase of the network easier. When modelling complex scenarios, the Bayesian Network representing the case of interest can become extremely large and so the inference of each posterior probability distribution can become an expensive and time consuming task, specially if any observation is introduced in the network. With this in mind, our goal was to create a library where the most recent techniques are exploited in order to reduce the complexity of calculations and thus the time and memory required for solving inference in a network of any size, trying to completely exploit all the resources available in modern computers. We also made possible to learn the structure of a network, with the possibility of adding constraints that must be respected, and the prior probability distribution of each variable from an input data set and to save it for a later use. In addition, we provide support for measuring the influence that a parent node has on the node of interest through different metrics. 

Relators:  Alessandro Savino, Stefano Di Carlo 
Academic year:  2021/22 
Publication type:  Electronic 
Number of Pages:  120 
Subjects:  
Corso di laurea:  Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering) 
Classe di laurea:  New organization > Master science > LM32  COMPUTER SYSTEMS ENGINEERING 
Aziende collaboratrici:  UNSPECIFIED 
URI:  http://webthesis.biblio.polito.it/id/eprint/22838 
Modify record (reserved for operators) 