Research Article
A New Evaluation Criteria for Learning Capability in OSA Context
@INPROCEEDINGS{10.1007/978-3-319-40352-6_1, author={Navikkumar Modi and Christophe Moy and Philippe Mary and Jacques Palicot}, title={A New Evaluation Criteria for Learning Capability in OSA Context}, proceedings={Cognitive Radio Oriented Wireless Networks. 11th International Conference, CROWNCOM 2016, Grenoble, France, May 30 - June 1, 2016, Proceedings}, proceedings_a={CROWNCOM}, year={2016}, month={6}, keywords={Cognitive radio Opportunistic spectrum access Reinforcement learning Multi-armed bandit Lempel-Ziv (LZ) complexity Optimal Arm Identification (OI) factor}, doi={10.1007/978-3-319-40352-6_1} }- Navikkumar Modi
Christophe Moy
Philippe Mary
Jacques Palicot
Year: 2016
A New Evaluation Criteria for Learning Capability in OSA Context
CROWNCOM
Springer
DOI: 10.1007/978-3-319-40352-6_1
Abstract
The activity pattern of different primary users (PUs) in the spectrum bands has a severe effect on the ability of the multi-armed bandit (MAB) policies to exploit spectrum opportunities. In order to apply MAB paradigm to opportunistic spectrum access (OSA), we must find out first whether the target channel set contains sufficient structure, over an appropriate time scale, to be identified by MAB policies. In this paper, we propose a criteria for analyzing suitability of MAB learning policies for OSA scenario. We propose a new criteria to evaluate the structure of random samples measured over time and referred as Optimal Arm Identification (OI) factor. OI factor refers to the difficulty associated with the identification of the optimal channel for opportunistic access. We found in particular that the ability of a secondary user to learn the activity of PUs spectrum is highly correlated to the OI factor but not really to the well known LZ complexity measure. Moreover, in case of very high OI factor, MAB policies achieve very little percentage of improvement compared to random channel selection (RCS) approach.