Learning (to improve) Optimal LP based planning heuristics
Title: Learning (to improve) Optimal LP based planning heuristics
SNIC Project: SNIC 2022/22-1074
Project Type: SNIC Small Compute
Principal Investigator: Paul Höft <paul.hoft@liu.se>
Affiliation: Linköpings universitet
Duration: 2022-11-07 – 2023-12-01
Classification: 10201
Keywords:

Abstract

We plan to use the resources to explore and classify the generalization capabilities of Linear Programm based planning heuristics over systematically generated pattern heuristics. Previous results of ours show that heuristics based on linear programs have an exploitable generalization. Our goal is to use Machine Learning to predict the reusability of these heuristics to speed up the computation of existing classical planning heuristics. We further plan to extend that approach to reinforcement learning techniques.