A Recursive Random Search Algorithm for Optimizing Network Protocol Parameters

The performance of several network protocols can be significantly enhanced by tuning their parameters. The optimization of network protocol parameters can be modeled as a “black-box” optimization problem with unknown, multi-modal and noisy objective functions. In this paper, a recursive random search algorithm is proposed to address this type of optimization problems. The new algorithm takes advantage of the favorable statistical properties of random sampling and achieves high efficiency without imposing extra restrictions, e.g., differentiability, on the objective function. It is also robust to noise in the evaluation of objective function since it use no traditional noise-susceptible local search techniques. The proposed algorithm is tested on classical benchmark functions and its performance compared with a multi-start hillclimbing algorithm. The algorithm is also integrated with a new on-line simulation system which attempts to automate network management by tuning protocol parameters when network conditions change significantly. We present the application of this on-line simulation system in enhancing the performance of network protocols, such as, RED and OSPF.