Black-box;
Expensive (i.e. taking a few minute to perform one evaluation);
Rugged and sparse landscape with numerous local optima;
There exist other performance attributes (e.g. throughput in addition to latency).
Landscape of the objective latency with respect to Splitters and Counters (out of the six configuration options)
Given the characteristics of the problem, we use an additional performance attribute to construct an artificial bi-objective optimisation model to help the search jump out of local optima.
Experiments show the effectiveness of our model - up to 42% gain, while using as low as 24% computational resources than state of the arts.