A Unified Search Framework for Large-scale Black-box Optimization

The parameter configuration of a network protocol can be formulated as a black-box optimization problem with network simulation evaluating the performance of the blackbox, i.e., the network. This paper proposes a unified search framework (USF) to handle such large-scale black-box optimization problems. The framework is designed to provides a general platform on which tailored optimization algorithms can be constructed easily for various types of problems. Therefore, it can be applied to the configuration of different network protocols. In the USF, various samplers are provided as basic building blocks and each of them implements a certain search technique. For a specific problem, a selection of samplers can be used to construct an appropriate search algorithm. These samplers are run in parallel and coordinated with various type of memories which selectively store the samples generated by samplers. The USF also include a resource management mechanism, which can manage parallel computing devices, for example, a network of workstations, and allocate the available computing resources to samplers according to the predefined allocation strategy. The benchmark tests are presented in this paper to demonstrate the flexibility and advantages of the USF.