Affiliations: Electrical Engineering Department, The City College of New York, New York, NY 10031, USA, Tel.: +1 212 650 7248, Fax: +1 212 650 8249, E-mail: [email protected]
Note: [] With Lucent Technologies Bell Labs Innovations, Whippany, NJ, USA. Corresponding author.
Abstract: This paper presents a new approach to the problem of congestion control arising at the User-to-Network Interface (UNI) of the ATM-hased Broadband Integrated Services Digital Networks (B-ISDN). Our approach employs an adaptive rate-based feedback control algorithm using reinforcement learning Neural Networks (NNs). We show that this new approach is very effective in limiting the buffer's occupancy levels and hence, minimizing congestion episodes. Moreover, the statistical multiplexing gain is greatly enhanced as more sources can be supported over the same buffer. The reinforcement learning NN controller provides an adaptive optimal control solution. This is achieved via the formulation of a performance measure function (cost function) that is used to, adaptively, tune the weights of the NN. The cost function is defined in terms of two main objectives: 1) to minimize the Cell Loss Rate (CLR), i.e., control congestion, and 2) to preserve the quality of the voice/video traffic via maintaining the original coding rate of the multimedia sources. The output of the NN controller is fed-back to the input sources, to throttle their arrival rates via reducing the coding rate (i.e., decreasing the number of bits per sample). Simulation results show that the NN control system is adaptive in the sense that it is applicable to different types of multimedia traffic. Also, the control signal maximizes the performance of the system which is defined in terms of its performance measure function.
