A framework for constrained large-scale multi-objective white-box problems based on two-scale optimization through decision transfer

Abstract

Most existing constrained multi-objective evolutionary algorithms (CMOEAs) are not so efficient when handling constrained large-scale multi-objective problems (CLSMOPs). To overcome white-box CLSMOPs with definitive objective functions, a two-scale optimization framework based on decision transfer, which integrates dimensionality reduction of large-scale decision variables and constraint handling technology, is proposed. The Lagrange multiplier is first used to construct the two-scale optimization model, which bridges original large-scale decision space of variables and small-scale (2-scale) decision space of objective-constraint parameter. The decision transfer algorithm is then designed to switch between large-scale original decision space and small-scale parametric decision space, while achieving the maximum dimensionality reduction. Finally, the two-scale evolution strategy is proposed for the alternative optimizations in the two decision spaces, which emphasize objectives and constraints, respectively. In summary, the optimization in the large-scale space pushes the population to unconstrained Pareto front (PF), the optimization in the small-scale space helps the population cross the infeasible areas to approach constrained PF, and the offspring generation by Lagrange multiplier is beneficial to both objectives and constraints. Eight representative and state-of-the-art CMOEAs have been embedded into the CLDTEA framework to demonstrate its effectiveness through comparative experiments on CLSMOPs with equalityandinequalityconstraints and 1000 decision variables. Experimental results show that CLDTEA can significantly improve the performance of these basic CMOEAs.