Conservative and Greedy Approaches to Classification-based Policy Iteration

Mohammad Ghavamzadeh 1 Alessandro Lazaric 1
1 SEQUEL - Sequential Learning
LIFL - Laboratoire d'Informatique Fondamentale de Lille, LAGIS - Laboratoire d'Automatique, Génie Informatique et Signal, Inria Lille - Nord Europe
Abstract : The existing classification-based policy iteration (CBPI) algorithms can be divided into two categories: {\em direct policy iteration} (DPI) methods that directly assign the output of the classifier (the approximate greedy policy w.r.t.~the current policy) to the next policy, and {\em conservative policy iteration} (CPI) methods in which the new policy is a mixture distribution of the current policy and the output of the classifier. The conservative policy update gives CPI a desirable feature, namely the guarantee that the policies generated by this algorithm improve at each iteration. We provide a detailed algorithmic and theoretical comparison of these two classes of CBPI algorithms. Our results reveal that in order to achieve the same level of accuracy, CPI requires more iterations, and thus, more samples than the DPI algorithm. Furthermore, CPI may converge to suboptimal policies whose performance is not better than DPI's.
Document type :
Conference papers
Complete list of metadatas

Cited literature [9 references]  Display  Hide  Download

https://hal.inria.fr/hal-00772610
Contributor : Alessandro Lazaric <>
Submitted on : Thursday, January 10, 2013 - 6:12:36 PM
Last modification on : Thursday, February 21, 2019 - 10:52:49 AM
Long-term archiving on : Thursday, April 11, 2013 - 4:08:46 AM

File

Ghavamzadeh.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-00772610, version 1

Collections

Citation

Mohammad Ghavamzadeh, Alessandro Lazaric. Conservative and Greedy Approaches to Classification-based Policy Iteration. AAAI - 26th Conference on Artificial Intelligence, Jul 2012, Toronto, Canada. ⟨hal-00772610⟩

Share

Metrics

Record views

297

Files downloads

160