MULTI-AGENT SERVICE SELECTION IN COMPETITIVE RESOURCE- CONSTRAINED ENVIRONMENTS Doctoral Thesis

We study the problem of service selection in environments with limited service capacity, under the assumptions characteristic of future network environments, namely that of dynamism, heterogeneity, lack of centralized authority and the consequent need to view individual entities in the network as selfinterested individuals. To this aim, we propose a framework for describing service selection problems in resource-constrained environments. Utilizing the game-theoretic framework of congestion games, we analyse some of the selection problem properties and identify the class of service-oriented architectures where non-cooperative approach is efficient. Informed by this analysis, in the second part of the thesis, we propose a reinforcement-learning-based adaptive mechanism which consumers in a dynamic service-oriented network can use to optimize their service selection. We perform an extensive evaluation of the mechanism on a range of simulated service selection problems in order the understand the dependence of the mechanism performance on problem parameters. Both analytical and experimental results indicate that non-cooperative service selection is feasible in networks where demand for services does not (significantly) exceeds available capacity and/or where service performance degrades gracefully with increasing load.