Efficient solutions for hybrid Markov decision processes specified through probabilistic logic programming: deep neural nets vs symbolic reasoning

• Grant number: 17/21763-3
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate (Direct)
• Effective date (Start): June 01, 2018
• Effective date (End): May 31, 2019
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Leliane Nunes de Barros
• Grantee: Thiago Pereira Bueno
• Supervisor: Scott Patrick Sanner
• Host Institution: Instituto de Matemática e Estatística (IME). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Research place: University of Toronto (U of T), Canada
• Associated to the scholarship: 16/22900-1 - Markov decision processes specified by probabilistic logic programming: representation and solution, BP.DD

Abstract

This research internship aims at further developing the proponent's PhD thesis on "Markov Decision Processes specified through Probabilistic Logic Programming" by extending its representational formalism to the realm of mixed continuous and discrete planning problems, as well as by investigating efficient algorithmic solutions based on symbolic reasoning and deep neural models. Given recent advances in statistical optimization for multidimensional non-convex functions, we propose to combine ideas from probabilistic logic programming and probabilistic planning into a deep learning approach that allows to leverage an expressive and efficient framework to handle decision-theoretic sequential problems modeled as hybrid Markov Decision Processes (MDPs). The expected results of this research proposal shall include: (i) exact and approximate algorithms to solve hybrid MDPs specified through Probabilistic Logic Programming; (ii) a principled compilation process to translate hybrid MDPs into deep neural nets; and (iii) comparative analysis between symbolic methods and optimization-based approaches to solve hybrid MDPs. (AU)