Multiagent Resource Allocation in Service Networks

Publikations-Art
Dissertation
Autoren
Karänke, P.
Erscheinungsjahr
2014
Herausgeber
Universität Hohenheim
Abstract

The term service network (SN) denotes a network of software services in which complex software applications are provided to customers by aggregating multiple elementary services. These networks are based on the service-oriented computing (SOC) paradigm, which defines the fundamental technical concepts for software services over electronic networks, e.g., Web services and, most recently, Cloud services. For the provision of software services to customers, software service providers (SPs) have to allocate their scarce computational resources (i.e., hardware and software) of a certain quality to customer requests. The SOC paradigm facilitates interoperability over organizational boundaries by representing business relationships on the software system level. Composite software services aggregate multiple software services into software applications. This aggregation is denoted as service composition. The loose coupling of services leads to SNs as dynamic entities with changing interdependencies between services.
For composite software services, these dependencies exist across SN tiers; they result from the procurement of services, which are themselves utilized to produce additional services, and constitute a major problem for resource allocation in SNs. If these dependencies are not considered, the fulfillment of agreements may become unaccomplishable (overcommitment). Hence, the consideration of service dependencies is crucial for the allocation of service providers resources to fulfill customer requests in SNs.
However, existing resource allocation methods, which could consider these dependencies -- such as combinatorial auctions with a central auctioneer for the whole SN -- are not applicable, since there are no central coordinating entities in SNs. The application of an allocation mechanism that does not consider these dependencies might negatively affect the actual service delivery; results are penalty payments as well as a damage to the reputation of the providers.
This research is conducted in accordance to the design science paradigm in information system research. It is a problem-solving paradigm, which targets the construction and evaluation of IT artifacts. The objectives of this research are to develop and evaluate an allocation protocol, which can consider multi-tier service dependencies without the existence of central coordinating entities. Therefore, an interaction protocol engineering (IPE) perspective is applied to solve the problem of multi-tier dependencies in resource allocation. This approach provides a procedure model for designing interaction protocols for multiagent systems, and is closely related to the well-established area of communication protocol engineering.
Automated resource allocation in SNs is analyzed in this research by representing the actors as autonomous software agents in the software system. The actors delegate their objectives to their software agents, which conduct the negotiations for service provision on their behalf. Thus, these agents communicate concerning the resource allocation; in this process, the sequence of communication interactions is crucial to the problem addressed. Interaction protocols define a structured exchange of defined messages between agents; they facilitate agent conversations.
When multiple agents have to reach agreements by negotiation and bargaining, such as in case with allocating scarce resources, game theory provides means to formalize and analyze the most rational choice of actions for the interacting agents. Based on a formal framework for resource allocation in SNs, this research first performs a game-theoretic problem analysis; it is concerned with the existence, as well as the complexity of computing optimal allocations. In addition, Nash equilibria are analyzed for optimal allocations. Second, a distributed, auction-based allocation protocol, which prevents overcommitments and guarantees socially optimal allocations for single customer requests under certain assumptions, is proposed. Therefore, a game-theoretic model and an operationizable specification of the protocol are presented. Third, it is formally verified that the protocol enables multi-tier resource allocation and avoids overcommitments by proofs for the game-theoretic model and by model checking for the interaction protocol specification; using the model checker Spin, safety properties like the absence of deadlock are as well formally verified as the protocol enabling multi-tier resource allocation. Fourth, the efficacy and the benefits of the proposed protocol are demonstrated by multiagent simulation for concurrent customers. The experimental evaluation provides evidence of the protocols efficiency compared to the socially optimal allocation as a centralized benchmark in different settings, e.g., network topologies and different bidding policies.


Beteiligte Einrichtungen

Weitere Informationen