Software Service Selection by Multi-level Matching and Reinforcement Learning

The software realization of distributed systems is typically achieved as loose coalitions of independently created services. The selection of such services, to act as building blocks of a distributed system, is a critical task that requires discovery and matching activities. This selection task is generally based on simple matching techniques and without any notion of customization. This paper presents a method to achieve the service discovery process using the principles of multilevel matching based on multi-level specifications and customization based on reinforcement learning techniques. In this method, services are selected dynamically using an on-line performance-based reinforcement feedback. In contrast to methods which require the services to actually carry out a task before being selected, in the method proposed in this paper, service selection is carried out using only specification matching, thereby eliminating a large amount of redundant computation. Experimental results are presented in the context of a information classification system. These experiments demonstrate that a high degree of performance can be achieved at a much reduced computational cost using only multi-level specification-matching based reinforcement feedback signals.