ABSTRACT
It is reasonable to assume that the changing of the optimization environment is smooth when considering a dynamic multi-objective optimization problem. Learning techniques are widely used to explore the dependence structure to facilitate population re-initialization in evolutionary search paradigms. The aim of the learning techniques is to discover knowledge from history information, thereby to track the movement of the optimal front quickly through good initialization when a change occurs. In this article, a new learning strategy is proposed, where the main ideas are (1) to use mutual information to identify the relationship between previously found approximated solutions; (2) to use a stable matching mechanism strategy to associate previously found optimal solutions bijectively; and (3) to re-initialize the new population based on a kinematics model. Controlled experiments were carried out systematically on some widely used test problems. Comparison against several state-of-the-art dynamic multi-objective evolutionary algorithms showed comparable performance in favour of the developed algorithm.
Disclosure statement
No potential conflict of interest was reported by the authors.
Notes
1. CER predicts the location of a new solution by extrapolation of previous solutions in adjacent generations, i.e. , where i is the index in the population, t is the generation index, and
is defined as
and
with equal probability where
.
2. In FPS, an autoregressive model is learned from the previous 10 generations and used to predict the new solutions as follows: , where
are the coefficient matrices and p is the degree of the AR model. In the experiment, p=2 was taken, as suggested in Hatzakis and Wallace (Citation2006).