Fast Reinforcement Learning with Large Action Sets Using Error-Correcting Output Codes for MDP Factorization

Gabriel Dulac-Arnold 1 Ludovic Denoyer 1 Philippe Preux 2 Patrick Gallinari 1
1 MALIRE - Machine Learning and Information Retrieval
LIP6 - Laboratoire d'Informatique de Paris 6
2 SEQUEL - Sequential Learning
LIFL - Laboratoire d'Informatique Fondamentale de Lille, LAGIS - Laboratoire d'Automatique, Génie Informatique et Signal, Inria Lille - Nord Europe
Abstract : The use of Reinforcement Learning in real-world scenarios is strongly limited by issues of scale. Most RL learning algorithms are unable to deal with problems composed of hundreds or sometimes even dozens of possible actions, and therefore cannot be applied to many real-world problems. We consider the RL problem in the supervised classification framework where the optimal policy is obtained through a multiclass classifier, the set of classes being the set of actions of the problem. We introduce error-correcting output codes (ECOCs) in this setting and propose two new methods for reducing complexity when using rollouts-based approaches. The first method consists in using an ECOC-based classifier as the multiclass classifier, reducing the learning complexity from O(A2) to O(Alog(A)) . We then propose a novel method that profits from the ECOC's coding dictionary to split the initial MDP into O(log(A)) separate two-action MDPs. This second method reduces learning complexity even further, from O(A2) to O(log(A)) , thus rendering problems with large action sets tractable. We finish by experimentally demonstrating the advantages of our approach on a set of benchmark problems, both in speed and performance.
Document type :
Conference papers
European Conference on Machine Learning, Sep 2012, Bristol, United Kingdom. Springer, Machine Learning and Knowledge Discovery in Databases, 7524, pp.180-194, 2012, Lecture Notes in Computer Science. <http://link.springer.com/chapter/10.1007/978-3-642-33486-3_12>. <10.1007/978-3-642-33486-3_12>


https://hal.inria.fr/hal-00747729
Contributor : Preux Philippe <>
Submitted on : Thursday, November 8, 2012 - 3:21:30 PM
Last modification on : Saturday, December 5, 2015 - 2:03:10 AM
Document(s) archivé(s) le : Saturday, February 9, 2013 - 3:41:39 AM

File

version.officielle.Springer.pd...
Publisher files allowed on an open archive

Identifiers

Collections

Citation

Gabriel Dulac-Arnold, Ludovic Denoyer, Philippe Preux, Patrick Gallinari. Fast Reinforcement Learning with Large Action Sets Using Error-Correcting Output Codes for MDP Factorization. European Conference on Machine Learning, Sep 2012, Bristol, United Kingdom. Springer, Machine Learning and Knowledge Discovery in Databases, 7524, pp.180-194, 2012, Lecture Notes in Computer Science. <http://link.springer.com/chapter/10.1007/978-3-642-33486-3_12>. <10.1007/978-3-642-33486-3_12>. <hal-00747729>

Export

Share

Metrics

Record views

218

Document downloads

66