Four Multi-agent Architectures for Intelligent Network Load Management

Abstract

Four different multi-agent system architectures for a dynamic resource allocation problem are presented and evaluated. Although all the architectures use three types of agents, i.e., quantifiers that act on the behalf the providers, allocators that act on the behalf of the customers, and distributors that decide how the available resources should be divided between the customers, they differ with respect to the degree of distribution of control and the degree of synchronization. The architectures are evaluated through simulation experiments concerning load balancing and overload control of Intelligent Networks. A number of aspects are compared, e.g., how fast the system re-allocates the resources when there are sudden changes in demand, how well the load is balanced between the providers, how well the resources are utilized, how fast a customer gets response, how fairly the system treats the customers, how robust the system is, and the amount of communication overhead. Some of the conclusions are that the asynchronous architectures react faster and that the centralized architectures balance the load better.On the other hand, the centralized architectures have a single point of failure and the asynchronous architectures tend to use more communication overhead.