In Proc. of International Conference on Telecommunications (ICT), 1998.
We investigate the usage and functionality of an intelligent software agent at the user-side in an ATM communications environment where multiple users (not necessarily identical) are served by and charged for ABR connections. We formulate an optimization problem pertaining to the case where bandwidth is allocated to each user proportionally to the amount of money he is willing to pay per unit time. Each user is assumed to select his willingness-to-pay so as to maximize his net benefit, i.e. the difference of the utility induced to the user by the QoS received minus his willingness-to-pay. The optimal selection depends on the network state, namely on the total willingness-to-pay by other users and on the total capacity available for ABR connections. We develop an intelligent agent that replaces the user in choosing the willingness-to-pay, adaptively to the network state, by learning the user preferences. This is based on past history of user choices for a wide range of network states. We analyze the main issues related to the above problem, and we present an algorithm for the intelligent agent to learn user preferences and select the willingness-to-pay on his behalf. We also provide simulation results showing that the performance of this algorithm is very close to optimal. Our approach can be applied to more general cases of economic sharing of network resources, and offers new capabilities for resource management, while it is not specific to ABR. It can in general be applied to elastic services, such as TCP/IP.
Paper available as a [.ps.gz] file. Google Scholar citations.