Solving a markov decision process multidimensional problem with tensor decomposition

Markov Decision Process (MDP) is a model used for planning decision-making of agents in stochastic and completely observable environments. Although much research is focused on solving atomic MDP problems in tabular forms or MDP problems with factored representations, none is based on tensor decompos...

Full description

Bibliographic Details
Author: Kuinchtner, Daniela
Format: master thesis
Status:Published version
Publication Date:2021
Country:Brasil
Institution:Pontifícia Universidade Católica do Rio Grande do Sul (PUCRS)
Repository:Biblioteca Digital de Teses e Dissertações da PUC_RS
Language:English
OAI Identifier:oai:tede2.pucrs.br:tede/9832
Online Access:http://tede2.pucrs.br/tede2/handle/tede/9832
Access Level:Open access
Keyword:Artificial Intelligence
CANDECOMP-PARAFAC Decomposition
Markov Decision Process
Tensor Algebra
Tensor Decomposition
Inteligência Artificial
Decomposição CANDECOMP-PARAFAC
Processo de ´Decisão de Markov
Ágebra Tensorial
Decomposição de Tensores
CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
Description
Summary:Markov Decision Process (MDP) is a model used for planning decision-making of agents in stochastic and completely observable environments. Although much research is focused on solving atomic MDP problems in tabular forms or MDP problems with factored representations, none is based on tensor decomposition methods. Solving MDPs using tensor algebra offers the prospect of leveraging advances in tensor-based calculations to increase MDP solvers’ efficiency. In this research, first, we formalize MDP multidimensional problems using tensor algebra. Second, we develop an MDP solver using the CANDECOMP-PARAFAC tensor decomposition method to compact state transition matrices. The solver uses the value iteration and policy iteration algorithms to compute the solution compactly. Then, we empirically evaluate the compact algorithms compared to tabular methods. As a result, we show that the tensor approach can compute larger problems using substantially less memory, opening up new possibilities for tensor-based methods for stochastic planning.