Adaptive Intelligent Scheduling and Traffic Management for ATM Networks

High speed packet switched networks provide numerous challenges for machine learning based control. In this proposal, we address the problem of scheduling the transmission of packets in an ATM switch supporting multiple traffic types with different delay and loss requirements. We propose an adaptive transmission scheduling algorithm, called Urgency Scheduling, that combines elements of memory based function approximation, reinforcement learning and dynamic programming. The resulting control will be validated by testing it with realistic traffic followed by direct comparison with standard service disciplines. Specific Phase I tasks are: (i) Perform a State-Of-The-Art survey of adaptive scheduling algorithms for B-ISDN applications with a focus on the usefulness of machine learning in terms of cost effectiveness, performance, and implementation for cell scheduling in C3I based ATM networks. (ii) Reformulate the Urgency Scheduling algorithm to investigate specific features of the model, especially, the sensitivity of performance to model parameters. (iii) Run the simplified model using off-line or simulated ATM data. (iv) Identify the learned strategies and investigate their implementation in a fuzzy rule-based framework. (v) Prepare final report describing results, Phase II recommendations and applicability to C3I systems. This project will be a collaborative effort between Scientific Systems and GTE Government Systems. GTE will supplement the effort with internal R&D funds and provide additional consulting support from GTE Labs (Dr Sutton). GTE will also commercialize the R&D in Phase III.