This paper investigates Factored Markov Decision Processes with Imprecise Probabilities; that is, Markov Decision Processes where transition probabilities are imprecisely specified, and where their specification does not deal directly with states, but rather with factored representations of states. We first define a Factored MDPIP, based on a multilinear formulation for MDPIPs; then we propose a novel algorithm for generation of Gamma-maximin policies for Factored MDPIPs. We also developed a representation language for Factored MDPIPs (based on the standard PPDDL language); finally, we describe experiments with a problem of practical significance, the well-known System Administrator Planning problem.
Keywords. Imprecise Markov Decision Processes (MDPIPs), Probabilistic Planning and PPDDL, Knowledge Representation Languages, Multilinear programming.
Paper Download
The paper is availabe in the following formats:
Plenary talk : Press here to get the file of the presentation.
Poster : Press here to get the file of the poster.
Authors addresses:
Karina Delgado
Rua do Matao 1010
Cidade Universitaria Sao Paulo, SP
Leliane Barros
Departamento de Computacao
Instituto de Matematica e Estatistica
Cidade Universitaria, CEP 05508900
Sao Paulo, SP
Fabio Cozman
Av. Prof. Mello Moraes, 2231
Cidade Univesitaria, CEP 05508-900
Sao Paulo, SP - BRAZIL
Ricardo Shirota Filho
Laboratório de Tomada de Decisão
A/C Prof. Dr. Fabio G. Cozman
Escola Politécnica da USP
Av. Prof. Mello Moraes, 2231
CEP: 05356-000
São Paulo, SP, BRAZIL
E-mail addresses:
| Karina Delgado | kvd@ime.usp.br |
| Leliane Barros | leliane@ime.usp.br |
| Fabio Cozman | fgcozman@usp.br |
| Ricardo Shirota Filho | ricardo.shirota@poli.usp.br |