Selection of auxiliary objectives using landscape features and offline learned classifier

Abstract

In order to increase the performance of an evolutionary algorithm, additional auxiliary optimization objectives may be added. It is hard to predict which auxiliary objectives will be the most efficient at different stages of optimization. Thus, the problem of dynamic selection between auxiliary objectives appears. This paper proposes a new method for efficient selection of auxiliary objectives, which uses fitness landscape information and problem meta-features. An offline learned meta-classifier is used to dynamically predict the most efficient auxiliary objective during the main optimization run performed by an evolutionary algorithm. An empirical evaluation on two benchmark combinatorial optimization problems (Traveling Salesman and Job Shop Scheduling problems) shows that the proposed approach outperforms similar known methods of auxiliary objective selection.