Memory based self-adaptive sampling for noisy multi-objective optimization

The paper proposes a novel strategy to adapt sample size of population members of a multi-objective optimization (MOO) problem, where the objective surface is contaminated with noise. The sample size, used for periodic fitness evaluation of a solution, is adapted based on the fitness variance of a sub-population in its neighborhood and is referred to as local neighborhood fitness variance (LNFV). The constraint of selecting accurate functional relationship between sample size and LNFV is surmounted here by employing a novel memory-based sample size adaptation policy. In the early exploration phase of a MOO, the policy memorizes the success or failure of sample sizes assigned to solutions with specific LNFVs. These success and failure history are later utilized to guide solutions of future generations to carefully select sample sizes based on their individual LNFVs. Experiments undertaken disclose the superiority of the proposed realization to the existing counterparts and the state-of-the-art techniques. The proposed algorithms have also been applied on a multi-robot box-pushing problem where the sensory data of twin robots are contaminated with noise. Experimental results here too reveal the efficiency of the proposed realizations in terms of minimization of execution time and energy consumed by the twin robots. (C) 2019 Elsevier Inc. All rights reserved.