Model Free Approach to Adaptive Parameters Optimisation using Adaptive Decaying Merit Score (ADMS): A Preliminary Study

摘要

Adaptive parameter control is a proven method to increase efficiency of any optimisation algorithm and have has been universally proven on various benchmark problems. In this research report, a Universal model free method of adaptive control of optimisation parameters was proposed. The goal of the research is to explore the method that can be used for any model of optimisation algorithm (`Swiss army knife' approach). The proposed method utilises a score system in which points are rewarded for improvement found on individual agent performance while non-improving will experience a score-decay. The proposed method was tested on two types of common optimisation algorithm, Truncated Lěvy flight and the Differential Evolution (DE). This Adaptive parameter strategy is solely based on search agent's performance and hence is not algorithm specific (model free). As a preliminary study, this report will only focus on Lower Dimensional problems (D=10) using CEC 2014 benchmark test sets. The results showed that it work well on DE but gave mediocre results for Le checkvy Flight when compared with their static parameter counterparts. This showed performance varies across algorithms types but verified the adaptive capability of the proposed method.