A benchmark for Maximum-a-Posteriori Inference algorithms in discrete Sum-Product Networks

• Autor(a) principal: Ribeiro, Heitor Reis
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-19062021-063556/
• topic: Maximum-a-posteriori; Maximum-a-posteriori; Modelos probabilísticos; Probabilistic models; Redes soma-produto; Sum-product networks

The solution to Maximum-a-Posteriori Inference problems in Sum-Product Networks provides the most probable configuration of the Random Variables encoded in its structure; a key step in Probabilistic reasoning that can be used for many applications, such as image auto-completion. It has been proven that this problem is NP-Hard (even to approximate) in Sum-Product Networks. Multiple algorithms have been developed to reach either approximate or exact solutions to this problem, but the experiments have been limited. In this Dissertation, we provide descriptions, analysis, and a benchmark for experimental testing for algorithms that solve this problem. We conclude that, given limited time, a Local Search algorithm starting with a solution found by the Argmax-Product algorithm reaches, on average, better results on the tested datasets.