Thinking outside the TBox multiparty service matchmaking as information retrieval
Lambert, David James
Service oriented computing is crucial to a large and growing number of computational undertakings. Central to its approach are the open and network-accessible services provided by many different organisations, and which in turn enable the easy creation of composite workflows. This leads to an environment containing many thousands of services, in which a programmer or automated composition system must discover and select services appropriate for the task at hand. This discovery and selection process is known as matchmaking. Prior work in the field has conceived the problem as one of sufficiently describing individual services using formal, symbolic knowledge representation languages. We review the prior work, and present arguments for why it is optimistic to assume that this approach will be adequate by itself. With these issues in mind, we examine how, by reformulating the task and giving the matchmaker a record of prior service performance, we can alleviate some of the problems. Using two formalisms—the incidence calculus and the lightweight coordination calculus—along with algorithms inspired by information retrieval techniques, we evolve a series of simple matchmaking agents that learn from experience how to select those services which performed well in the past, while making minimal demands on the service users. We extend this mechanism to the overlooked case of matchmaking in workflows using multiple services, selecting groups of services known to inter-operate well. We examine the performance of such matchmakers in possible future services environments, and discuss issues in applying such techniques in large-scale deployments.