Confidence-Based Algorithm Parameter Tuning with Dynamic Resampling.

This work presents an algorithm for tuning the parameters of stochastic search heuristics, the Robust Parameter Searcher (RPS). RPS is based on the Nelder-Mead Simplex algorithm and on confidence-based comparison operators. Whilst the latter algorithm is known for its robustness under noise in objective function evaluation, the confidence-based comparison endows the tuning algorithm with additional resilience against the intrinsic stochasticity which exists in the evaluation of performance of stochastic search heuristics. The proposed methodology was used to tune a Differential Evolution strategy for optimizing real-valued functions, with a limited function evaluation budget. In the computational experiments, RPS performed significantly better than other well-known tuning strategies from the literature.