6th International Conference on Performance Evaluation Methodologies and Tools

Research Article

SLA negotiation: Experimental Observations on Learning Policies

Download588 downloads
  • @INPROCEEDINGS{10.4108/valuetools.2012.250344,
        author={Mohamed Lamine Lamali and Helia Pouyllau and Dominique Barth},
        title={SLA negotiation: Experimental Observations on Learning Policies},
        proceedings={6th International Conference on Performance Evaluation Methodologies and Tools},
        publisher={IEEE},
        proceedings_a={VALUETOOLS},
        year={2012},
        month={11},
        keywords={quality of service sla negotiation mixed nash equilibrium reinforcement learning},
        doi={10.4108/valuetools.2012.250344}
    }
    
  • Mohamed Lamine Lamali
    Helia Pouyllau
    Dominique Barth
    Year: 2012
    SLA negotiation: Experimental Observations on Learning Policies
    VALUETOOLS
    ICST
    DOI: 10.4108/valuetools.2012.250344
Mohamed Lamine Lamali1,*, Helia Pouyllau2, Dominique Barth3
  • 1: Alcatel-Lucent Bell Labs
  • 2: Alcatel-Lucent Bell Labs France
  • 3: Laboratoire PRiSM. University of Versailles
*Contact email: mohamed_lamine.lamali@alcatel-lucent.com

Abstract

The Internet has moved to content broadcasting and one might anticipate future evolutions of the supported applications. Meanwhile, the Internet business model remains the same from the early-days. While, on the technological side, many discussions assess the ossification of the Internet, the Internet traded good is still reachability. Some authors argue that the technical ossification is a consequence of the economic one. But adopting a clean-slate economic model is as challenging as adopting a clean-slate architecture. The new system must meet requirements on profitability and stability while tackling complex issues. In this paper, we focus on the proposal of enriching Service Level Agreements (SLAs), which are contracts among Network Service Providers (NSPs) with Quality of Service (QoS) information. We propose a game model of the SLA negotiation among NSPs in order to study how some learning algorithms converge to stable conditions, which are mixed Nash Equilibria in this case. Computing a mixed Nash equilibrium is PPAD-complete; the corresponding algorithms are thus quite complex. In previous works, some authors studied the convergence of Reinforcement Learning techniques to pure and mixed Nash Equilibria. Learning mixed Nash Equilibria seems harder. Hence, we rather experimentally observe how such algorithms can, according to different policies, converge to mixed Nash Equilibria, and also how profitable they are for the NSPs.