A Constrained Multi-objective Optimization Algorithm Based On Backwards Guided Extreme Learning Machine

Evolutionary algorithms simulate the evolutionary process of the natural world,adapting to the external environment through genes and mutations to achieve superiority and evolution from generation to generation.However,in the past many years,there has been no great breakthrough in the way of generating offspring of evolutionary algorithms,which is still dominated by random generation of offspring and lacks an effective guidance mechanism.This makes it difficult to have an effective correspondence between the generation of the solutions of the offspring in the decision space and the solutions in the objective space corresponding to the calculation of the offspring by the real function afterwards,thus severing the connection between the objective space and the decision space.That is,solutions in the objective space can only be used for environmental selection,while decision space solutions can only be used for evolution.This lack of guidance in the transformation relationship leads to the evolution of the solution in the decision space is full of randomness,and there is no guarantee that the performance of the offspring in the objective space will be optimized.In this paper,we investigate on the constrained multi-objective problem,and our main work is as follows.I.A backward guidance strategy is proposed,which aims to guide the decision space population evolution by backward prediction of the objective vector,so that the decision space optimization problem is transformed into an objective space optimization problem and can provide a better guidance for population evolution.II.The backward guidance strategy is combined with MOEA/D-DE algorithm and infeasible solution archiving method to form an algorithm ORGA for solving constrained optimization problems,and experiments are conducted with other four classical algorithms on CTP series test problems and TYPE series test problems,respectively.III.The backward guidance strategy is combined with the NSGA-II-CDP algorithm and the gray prediction method to form an algorithm GPRGA for solving dynamic constrained multi-objective optimization problems,and experiments are conducted on a variant of the DTYPE series of problems modified from the TYPE problem with the other two seed generation strategies based on the same algorithm NSGA-II-CDP.The experimental comparison shows that both ORGA and GPRGA exhibit convergence ability and stability due to similar algorithms in the corresponding experiments.This also illustrates the effectiveness of the backward guidance strategy on difficult problems from the side.In addition,the ability to combine with different algorithms conveniently also shows that the backward guidance strategy has the advantage of good reusability.