Unsupervised Learning of Policy Sketches for Classical Planning
Title: Unsupervised Learning of Policy Sketches for Classical Planning
SNIC Project: SNIC 2021/22-595
Project Type: SNIC Small Compute
Principal Investigator: Dominik Drexler <dominik.drexler@liu.se>
Affiliation: Linköpings universitet
Duration: 2021-08-09 – 2022-09-01
Classification: 10201
Homepage: https://rleap-project.github.io/
Keywords:

Abstract

In this work, we are trying to learn policy sketches for solving tractable classical planning domains in provably low polynomial time. A policy sketch R decomposes a problem into subproblems. The challenge is to automatically learn a suitable sketch that works on a large class of problems over a common domain. There are certain properties that a sketch has to satisfy such as being well-formed and has bounded and small sketch width in order to be suitable. We use answer set programming for finding suitable sketches automatically.