Reinforcement Learning Techniques for Decentralized Self-adaptive Service Assembly
This paper proposes a self-organizing fully decentralized solution for the service assembly problem, whose goal is to guarantee a good overall quality for the delivered services, ensuring at the same time fairness among the participating peers. The main f
- PDF / 624,002 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 15 Downloads / 216 Views
Linnaeus University, V¨ axj¨ o, Sweden {mauro.caporuscio,mirko.dangelo}@lnu.se 2 Universit` a di Roma Tor Vergata, Rome, Italy [email protected] 3 Politecnico di Milano, Milan, Italy [email protected]
Abstract. This paper proposes a self-organizing fully decentralized solution for the service assembly problem, whose goal is to guarantee a good overall quality for the delivered services, ensuring at the same time fairness among the participating peers. The main features of our solution are: (i) the use of a gossip protocol to support decentralized information dissemination and decision making, and (ii) the use of a reinforcement learning approach to make each peer able to learn from its experience the service selection rule to be followed, thus overcoming the lack of global knowledge. Besides, we explicitly take into account load-dependent quality attributes, which lead to the definition of a service selection rule that drives the system away from overloading conditions that could adversely affect quality and fairness. Simulation experiments show that our solution self-adapts to occurring variations by quickly converging to viable assemblies maintaining the specified quality and fairness objectives.
1
Introduction
We consider a distributed peer-to-peer scenario, where a large set of peers cooperatively work to accomplish specific tasks. In general, each peer possesses the know-how to perform some tasks (offered services), but could require services offered by other peers to carry out these tasks. Scenarios of this type can be typically encountered in pervasive computing application domains like ambient intelligence or smart transportation systems, where several (from tens to thousands) services cooperate to achieve some common objectives [13]. A basic functional requirement for this scenario is to match required and provided services, so that the resulting assembly makes each peer able to correctly deliver its service(s). Besides this functional requirement, we also assume the existence of non functional requirements concerning the quality of the delivered service, expressed in terms of several quality attributes referring to different quality domains (e.g., performance, dependability, cost). c IFIP International Federation for Information Processing 2016 Published by Springer International Publishing Switzerland 2016. All Rights Reserved M. Aiello et al. (Eds.): ESOCC 2016, LNCS 9846, pp. 53–68, 2016. DOI: 10.1007/978-3-319-44482-6 4
54
M. Caporuscio et al.
Our goal is to devise a self-assembly procedure among the peers, aimed at fulfilling both functional and non functional requirements. For the latter, we aim in particular to maximize some notion of global utility expressed in terms of the quality attributes of the services delivered by peers in the system, ensuring at the same time fairness (i.e., no peer should be excessively favored or penalized with respect to others). Challenges to be tackled to achieve this goal include: (1) the presence of several functionally equivalent services, with di
Data Loading...
