Imperial College London
Portrait Picture

ProfessorAlessandraRusso

Faculty of Engineering, Department of Computing

Professor in Applied Computational Logic
 
 
 
//

Contact

 

+44 (0)20 7594 8312a.russo Website

 
 
//

Location

 

560Huxley BuildingSouth Kensington Campus

//

Summary

 

Publications

Citation

BibTex format

@inproceedings{Gomoluch:2019,
author = {Gomoluch, P and Alrajeh, D and Russo, A},
pages = {637--645},
publisher = {AAAI},
title = {Learning classical planning strategies with policy gradient},
url = {http://hdl.handle.net/10044/1/70378},
year = {2019}
}
Download

RIS format (EndNote, RefMan)

TY  - CPAPER
AB  - A common paradigm in classical planning is heuristic forwardsearch. Forward search planners often rely on simplebest-first search which remains fixed throughout the searchprocess. In this paper, we introduce a novel search frameworkcapable of alternating between several forward searchapproaches while solving a particular planning problem. Selectionof the approach is performed using a trainable stochasticpolicy, mapping the state of the search to a probability distributionover the approaches. This enables using policy gradientto learn search strategies tailored to a specific distributionsof planning problems and a selected performance metric,e.g. the IPC score. We instantiate the framework by constructinga policy space consisting of five search approachesand a two-dimensional representation of the planner’s state.Then, we train the system on randomly generated problemsfrom five IPC domains using three different performance metrics.Our experimental results show that the learner is ableto discover domain-specific search strategies, improving theplanner’s performance relative to the baselines of plain bestfirstsearch and a uniform policy.
AU  - Gomoluch,P
AU  - Alrajeh,D
AU  - Russo,A
EP  - 645
PB  - AAAI
PY  - 2019///
SN  - 2334-0843
SP  - 637
TI  - Learning classical planning strategies with policy gradient
UR  - http://hdl.handle.net/10044/1/70378
ER  -
Download