Self-Adaptive Sampling in Noisy Multi-objective Optimization
Pratyusha Rakshit, Amit Konar, Atulya K. Nagar · 2018
The paper proposes a novel sampling strategy to adapt sample size for periodic fitness evaluation of solutions of a multi-objective optimization (MOO) problem in the presence of noise in the objective surfaces. The existing works consider a fixed functional form of relationship between sample size of a solution and its local neighborhood fitness variance (LNFV). The non-decreasing monotonicity of the functional form being subjective to the noise characteristics, any fixed functional form is ineffective to allocate accurate sample size to a solution for all possible known/unknown noise distribution. This stalemate is overcome here by employing a novel learning induced sample size adaptation policy. The policy learns the success or failure of sample sizes assigned to solutions with specific LNFVs in the early exploration phase of an MOO and later utilizes the acquired knowledge to guide selection of sample size by solutions of future generations. Experiments undertaken reveal a statistically significant superiority of the proposed realizations to their existing counterparts with respect to inverted generational distance and hypervolume ratio metrics.