Elite Accumulative Sampling Strategies for Noisy Multi-Objective Optimisation

Abstract

When designing evolutionary algorithms one of the key concerns is the balance between expending function evaluations on exploration versus exploitation. When the optimisation problem experiences observational noise, there is also a trade-off with respect to accuracy refinement – as improving the estimate of a design’s performance typically is at the cost of additional function reevaluations. Empirically the most effective resampling approach developed so far is accumulative resampling of the elite set. In this approach elite members are regularly reevaluated, meaning they progressively accumulate reevaluations over time. This results in their approximated objective values having greater fidelity, meaning non-dominated solutions are more likely to be correctly identified. Here we examine four different approaches to accumulative resampling of elite members, embedded within a differential evolution algorithm. Comparing results on 40 variants of the unconstrained IEEE CEC’09 multi-objective test problems, we find that at low noise levels a low fixed resample rate is usually sufficient, however for larger noise magnitudes progressively raising the number of minimum resamples of elite members based on detecting estimated front oscillation tends to improve performance.