Minimum Regret Approach to Network Management under Uncertainty with Applications to Connection Admission Control and Routing

Abstract

This paper proposes a framework for network management intended to balance network performance under normal steady operational conditions with robustness under non-steady, and/or adverse conditions. Working in conjunction with anomaly and/or intrusion detection, the proposed framework allows the network to develop a set of measured responses to possible anomalies and external threats by minimizing the average maximum network losses, i.e., regrets or risks, due to uncertainty. Loss maximization guards against uncertainty within each scenario. Averaging of the maximum losses reflects an available information on the likelihood of different scenarios. The proposed framework includes Bayesian and minimax approaches as particular cases. The paper demonstrates how the proposed framework can alleviate high sensitivity of a cost-based admission and routing scheme to uncertain resource costs and possible presence of excessive and/or adversarial traffic. Specific examples include competitive and minimum interference admission and routing schemes.